server : handle If-None-Match weak ETags (#23916)

* vocab : add jina-embeddings-v2-base-zh (whitespace tokenizer)

* lowercase defaults to true

* type fix

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

.devops/zendnn.Dockerfile

@@ -0,0 +1,101 @@
+ARG UBUNTU_VERSION=24.04
+ARG BUILD_DATE=N/A
+ARG APP_VERSION=N/A
+ARG APP_REVISION=N/A
+
+FROM ubuntu:$UBUNTU_VERSION AS build
+
+RUN apt-get update && \
+    apt-get install -y gcc-13 g++-13 build-essential git cmake libssl-dev libomp-dev libnuma-dev python3 ca-certificates
+
+ENV CC=gcc-13 CXX=g++-13
+
+WORKDIR /app
+
+COPY . .
+
+RUN cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_ZENDNN=ON && \
+    cmake --build build -j $(nproc)
+
+RUN mkdir -p /app/lib && \
+    find build -name "*.so*" -exec cp -P {} /app/lib \;
+
+RUN mkdir -p /app/full \
+    && cp build/bin/* /app/full \
+    && cp *.py /app/full \
+    && cp -r conversion /app/full \
+    && cp -r gguf-py /app/full \
+    && cp -r requirements /app/full \
+    && cp requirements.txt /app/full \
+    && cp .devops/tools.sh /app/full/tools.sh
+
+## Base image
+FROM ubuntu:$UBUNTU_VERSION AS base
+
+ARG BUILD_DATE=N/A
+ARG APP_VERSION=N/A
+ARG APP_REVISION=N/A
+ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
+ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
+LABEL org.opencontainers.image.created=$BUILD_DATE \
+      org.opencontainers.image.version=$APP_VERSION \
+      org.opencontainers.image.revision=$APP_REVISION \
+      org.opencontainers.image.title="llama.cpp" \
+      org.opencontainers.image.description="LLM inference in C/C++" \
+      org.opencontainers.image.url=$IMAGE_URL \
+      org.opencontainers.image.source=$IMAGE_SOURCE
+
+RUN apt-get update \
+    && apt-get install -y libgomp1 libnuma1 curl \
+    && apt autoremove -y \
+    && apt clean -y \
+    && rm -rf /tmp/* /var/tmp/* \
+    && find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete \
+    && find /var/cache -type f -delete
+
+COPY --from=build /app/lib/ /app
+
+### Full
+FROM base AS full
+
+COPY --from=build /app/full /app
+
+WORKDIR /app
+
+RUN apt-get update \
+    && apt-get install -y \
+    git \
+    python3 \
+    python3-pip \
+    python3-wheel \
+    && pip install --break-system-packages --upgrade setuptools \
+    && pip install --break-system-packages -r requirements.txt \
+    && apt autoremove -y \
+    && apt clean -y \
+    && rm -rf /tmp/* /var/tmp/* \
+    && find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete \
+    && find /var/cache -type f -delete
+
+ENTRYPOINT ["/app/tools.sh"]
+
+### Light, CLI only
+FROM base AS light
+
+COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+
+WORKDIR /app
+
+ENTRYPOINT [ "/app/llama-cli" ]
+
+### Server, Server only
+FROM base AS server
+
+ENV LLAMA_ARG_HOST=0.0.0.0
+
+COPY --from=build /app/full/llama-server /app
+
+WORKDIR /app
+
+HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
+
+ENTRYPOINT [ "/app/llama-server" ]

.github/actions/ccache-clear/action.yml

@@ -0,0 +1,22 @@
+name: "ccache-clear"
+description: "Delete all GitHub Actions caches matching a key prefix"
+inputs:
+  key:
+    description: "Cache key prefix to match and delete"
+    required: true
+
+runs:
+  using: "composite"
+  steps:
+    - name: Clear caches
+      shell: bash
+      run: |
+        CACHES=$(gh cache list --key "ccache-${{ inputs.key }}" --json id,key --jq '.[] | "\(.id) \(.key)"' 2>/dev/null)
+        if [ -z "$CACHES" ]; then
+          echo "No caches found with key prefix: ${{ inputs.key }}"
+          exit 0
+        fi
+        while read -r id key; do
+          echo "Deleting cache: $id ($key)"
+          gh cache delete "$id"
+        done <<< "$CACHES"

.github/actions/windows-setup-cuda/action.yml

@@ -96,3 +96,34 @@ runs:
           echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
           echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
           echo "CUDA_PATH_V13_1=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+
+    - name: Install Cuda Toolkit 13.3
+      if: ${{ inputs.cuda_version == '13.3' }}
+      shell: pwsh
+      run: |
+          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3"
+          choco install unzip -y
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_crt/windows-x86_64/cuda_crt-windows-x86_64-13.3.33-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-13.3.29-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-13.3.33-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-13.3.33-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-x86_64/libcublas-windows-x86_64-13.5.1.27-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libnvvm/windows-x86_64/libnvvm-windows-x86_64-13.3.33-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-13.3.29-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_profiler_api/windows-x86_64/cuda_profiler_api-windows-x86_64-13.3.27-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-13.3.27-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cccl/windows-x86_64/cccl-windows-x86_64-13.3.3.3.1-archive.zip"
+          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3"
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\cuda_crt-windows-x86_64-13.3.33-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\cuda_cudart-windows-x86_64-13.3.29-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\cuda_nvcc-windows-x86_64-13.3.33-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\cuda_nvrtc-windows-x86_64-13.3.33-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\libcublas-windows-x86_64-13.5.1.27-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\libnvvm-windows-x86_64-13.3.33-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\cuda_nvtx-windows-x86_64-13.3.29-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\cuda_profiler_api-windows-x86_64-13.3.27-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\visual_studio_integration-windows-x86_64-13.3.27-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\cccl-windows-x86_64-13.3.3.3.1-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" /E /I /H /Y
+          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          echo "CUDA_PATH_V13_3=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.3" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8

.github/workflows/build-3rd-party.yml

@@ -22,9 +22,9 @@ concurrency:
 env:
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
-  LLAMA_LOG_COLORS: 1
-  LLAMA_LOG_PREFIX: 1
-  LLAMA_LOG_TIMESTAMPS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
 
 jobs:
   ubuntu-24-llguidance:

.github/workflows/build-and-test-snapdragon.yml

@@ -31,7 +31,7 @@ jobs:
   android-ndk-snapdragon:
     runs-on: ubuntu-latest
     container:
-      image: 'ghcr.io/snapdragon-toolchain/arm64-android:v0.6'
+      image: 'ghcr.io/snapdragon-toolchain/arm64-android:v0.7'
     defaults:
       run:
         shell: bash
@@ -61,7 +61,7 @@ jobs:
   linux-iot-snapdragon:
     runs-on: ubuntu-latest
     container:
-      image: 'ghcr.io/snapdragon-toolchain/arm64-linux:v0.6'
+      image: 'ghcr.io/snapdragon-toolchain/arm64-linux:v0.7'
     defaults:
       run:
         shell: bash

.github/workflows/build-android.yml

@@ -27,12 +27,12 @@ concurrency:
 env:
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
-  LLAMA_LOG_COLORS: 1
-  LLAMA_LOG_PREFIX: 1
-  LLAMA_LOG_TIMESTAMPS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
 
 jobs:
-  android:
+  default:
     runs-on: ubuntu-latest
 
     steps:
@@ -58,7 +58,7 @@ jobs:
           cd examples/llama.android
           ./gradlew build --no-daemon
 
-  android-ndk:
+  ndk:
     runs-on: ubuntu-latest
     container:
       image: 'ghcr.io/snapdragon-toolchain/arm64-android:v0.3'
@@ -91,3 +91,59 @@ jobs:
         with:
           name: llama-cpp-android-arm64-cpu
           path: pkg-adb/llama.cpp
+
+  arm64:
+    runs-on: ubuntu-latest
+
+    env:
+      NDK_VERSION: "29.0.14206865"
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      # note : disabled to spare some cache space (https://github.com/ggml-org/llama.cpp/pull/23789)
+      #        for some reason, the ccache does not improve the build time in this case
+      # example:
+      #   cache off: https://github.com/ggerganov/tmp2/actions/runs/26534713799/job/78160400831
+      #   cache on:  https://github.com/ggerganov/tmp2/actions/runs/26534713799/job/78224189394
+      #
+      #- name: ccache
+      #  uses: ggml-org/ccache-action@v1.2.21
+      #  with:
+      #    key: android-ubuntu-arm64
+      #    evict-old-files: 1d
+      #    save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Set up JDK
+        uses: actions/setup-java@v5
+        with:
+          java-version: 17
+          distribution: temurin
+
+      - name: Setup Android SDK
+        uses: android-actions/setup-android@40fd30fb8d7440372e1316f5d1809ec01dcd3699 # v4.0.1
+        with:
+          log-accepted-android-sdk-licenses: false
+
+      - name: Install NDK
+        run: |
+          sdkmanager "ndk;${{ env.NDK_VERSION }}"
+          echo "ANDROID_NDK=${ANDROID_SDK_ROOT}/ndk/${{ env.NDK_VERSION }}" >> $GITHUB_ENV
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build \
+            -DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK}/build/cmake/android.toolchain.cmake \
+            -DANDROID_ABI=arm64-v8a \
+            -DANDROID_PLATFORM=android-28 \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_NATIVE=OFF \
+            -DGGML_CPU_ALL_VARIANTS=ON \
+            -DGGML_OPENMP=OFF \
+            -DLLAMA_BUILD_BORINGSSL=ON \
+            -DGGML_RPC=ON
+          time cmake --build build --config Release -j $(nproc)

.github/workflows/build-apple.yml

@@ -32,12 +32,12 @@ concurrency:
 env:
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
-  LLAMA_LOG_COLORS: 1
-  LLAMA_LOG_PREFIX: 1
-  LLAMA_LOG_TIMESTAMPS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
 
 jobs:
-  macOS-latest-ios:
+  macos-latest-arm64:
     runs-on: macos-latest
 
     steps:
@@ -48,7 +48,80 @@ jobs:
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
         with:
-          key: macOS-latest-ios
+          key: apple-arm64
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          cmake -B build \
+            -DCMAKE_BUILD_RPATH="@loader_path" \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DLLAMA_BUILD_BORINGSSL=ON \
+            -DGGML_METAL_USE_BF16=ON \
+            -DGGML_METAL_EMBED_LIBRARY=OFF \
+            -DGGML_METAL_SHADER_DEBUG=ON \
+            -DGGML_RPC=ON
+          time cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
+          leaks -atExit -- ./build/bin/test-thread-safety -hf ggml-org/gemma-3-270m-qat-GGUF -ngl 99 -p "$(printf 'hello %.0s' {1..128})" -n 16 -c 512 -ub 32 -np 2 -t 2 -lv 1
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main -E "test-llama-archs" --verbose --timeout 900
+
+  macos-latest-x64:
+    runs-on: macos-15-intel
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: apple-x64
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          # Metal is disabled due to intermittent failures with Github runners not having a GPU:
+          # https://github.com/ggml-org/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
+          cmake -B build \
+            -DCMAKE_BUILD_RPATH="@loader_path" \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DLLAMA_BUILD_BORINGSSL=ON \
+            -DGGML_METAL=OFF \
+            -DGGML_RPC=ON \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3
+          time cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
+  macos-latest-ios:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      # TODO: this likely does not do anything - if yes, remove it
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: apple-ios
           evict-old-files: 1d
           save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
@@ -117,7 +190,7 @@ jobs:
           xcodebuild -downloadPlatform iOS
           xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' FRAMEWORK_FOLDER_PATH=./build-ios build
 
-  macOS-latest-tvos:
+  macos-latest-tvos:
     runs-on: macos-latest
 
     steps:
@@ -125,10 +198,11 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
+      # TODO: this likely does not do anything - if yes, remove it
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
         with:
-          key: macOS-latest-tvos
+          key: apple-tvos
           evict-old-files: 1d
           save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
@@ -150,7 +224,7 @@ jobs:
             -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
 
-  macOS-latest-visionos:
+  macos-latest-visionos:
     runs-on: macos-latest
 
     steps:
@@ -158,6 +232,14 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
+      # TODO: this likely does not do anything - if yes, remove it
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: apple-visionos
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
       - name: Build
         id: cmake_build
         run: |
@@ -176,7 +258,7 @@ jobs:
             -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
 
-  macOS-latest-swift:
+  macos-latest-swift:
     runs-on: macos-latest
     needs: macos-latest-ios-xcode
 
@@ -189,10 +271,11 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
+      # TODO: this likely does not do anything - if yes, remove it
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
         with:
-          key: macOS-latest-swift
+          key: apple-swift
           evict-old-files: 1d
           save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 

.github/workflows/build-cache.yml

@@ -28,7 +28,7 @@ jobs:
         id: cache-sdk
         with:
           path: ./vulkan_sdk
-          key: vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}
+          key: cache-gha-vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}
 
       - name: Setup Vulkan SDK
         if: steps.cache-sdk.outputs.cache-hit != 'true'
@@ -54,7 +54,7 @@ jobs:
   #      id: cache-toolchain
   #      with:
   #        path: ./spacemit_toolchain
-  #        key: spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}-${{ runner.os }}
+  #        key: cache-gha-spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}-${{ runner.os }}
 
   #    - name: Setup SpacemiT Toolchain
   #      if: steps.cache-toolchain.outputs.cache-hit != 'true'
@@ -81,7 +81,7 @@ jobs:
         id: cache-openvino
         with:
           path: ./openvino_toolkit
-          key: openvino-toolkit-v${{ env.OPENVINO_VERSION_FULL }}-${{ runner.os }}
+          key: cache-gha-openvino-toolkit-v${{ env.OPENVINO_VERSION_FULL }}-${{ runner.os }}
 
       - name: Setup OpenVINO Toolkit
         if: steps.cache-openvino.outputs.cache-hit != 'true'
@@ -108,7 +108,7 @@ jobs:
         id: cache-rocm
         with:
           path: C:\Program Files\AMD\ROCm
-          key: rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
+          key: cache-gha-rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
 
       - name: Setup ROCm
         if: steps.cache-rocm.outputs.cache-hit != 'true'

.github/workflows/build-cann.yml

@@ -29,74 +29,76 @@ concurrency:
 env:
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
-  LLAMA_LOG_COLORS: 1
-  LLAMA_LOG_PREFIX: 1
-  LLAMA_LOG_TIMESTAMPS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
 
 jobs:
-  openEuler-latest-cann:
-    defaults:
-      run:
-        shell: bash -el {0}
-    strategy:
-      matrix:
-        arch: [x86, aarch64]
-        chip_type: ['910b', '310p']
-        build: ['Release']
-        use_acl_graph: ['on', 'off']
-        exclude:
-          # 310P does not support USE_ACL_GRAPH=on
-          - chip_type: '310p'
-            use_acl_graph: 'on'
-    runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v6
-        with:
-          fetch-depth: 0
-
-      - name: Free up disk space
-        uses: ggml-org/free-disk-space@v1.3.1
-        with:
-          tool-cache: true
-
-      - name: Set container image
-        id: cann-image
-        run: |
-          image="ascendai/cann:${{ matrix.chip_type == '910b' &&  '8.5.0-910b-openeuler24.03-py3.11' || '8.5.0-310p-openeuler24.03-py3.11' }}"
-          echo "image=${image}" >> "${GITHUB_OUTPUT}"
-
-      - name: Pull container image
-        run: docker pull "${{ steps.cann-image.outputs.image }}"
-
-      - name: Build
-        env:
-          BUILD_TYPE: ${{ matrix.build }}
-          SOC_TYPE: ascend${{ matrix.chip_type }}
-          USE_ACL_GRAPH: ${{ matrix.use_acl_graph }}
-        run: |
-          HOST_UID=$(id -u)
-          HOST_GID=$(id -g)
-
-          docker run --rm \
-            -v "${PWD}:/workspace" \
-            -w /workspace \
-            -e SOC_TYPE=${SOC_TYPE} \
-            -e BUILD_TYPE=${BUILD_TYPE} \
-            -e USE_ACL_GRAPH=${USE_ACL_GRAPH} \
-            "${{ steps.cann-image.outputs.image }}" \
-            bash -lc '
-              set -e
-              yum install -y --setopt=install_weak_deps=False --setopt=tsflags=nodocs git gcc gcc-c++ make cmake openssl-devel
-              yum clean all && rm -rf /var/cache/yum
-              git config --global --add safe.directory "/workspace"
-              export LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/$(uname -m)-linux/devlib/:${LD_LIBRARY_PATH}
-              cmake -S . -B build \
-                  -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
-                  -DGGML_CANN=on \
-                  -DSOC_TYPE=${SOC_TYPE} \
-                  -DUSE_ACL_GRAPH=${USE_ACL_GRAPH}
-              cmake --build build -j $(nproc)
-
-              chown -R '"${HOST_UID}"':'"${HOST_GID}"' /workspace/build
-            '
+# TODO: this build is disabled to save Github Actions resources (https://github.com/ggml-org/llama.cpp/pull/23705)
+#       in order to enable it again, we have to provision dedicated runners  to run it
+#  openEuler-latest-cann:
+#    defaults:
+#      run:
+#        shell: bash -el {0}
+#    strategy:
+#      matrix:
+#        arch: [x86, aarch64]
+#        chip_type: ['910b', '310p']
+#        build: ['Release']
+#        use_acl_graph: ['on', 'off']
+#        exclude:
+#          # 310P does not support USE_ACL_GRAPH=on
+#          - chip_type: '310p'
+#            use_acl_graph: 'on'
+#    runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
+#    steps:
+#      - name: Checkout
+#        uses: actions/checkout@v6
+#        with:
+#          fetch-depth: 0
+#
+#      - name: Free up disk space
+#        uses: ggml-org/free-disk-space@v1.3.1
+#        with:
+#          tool-cache: true
+#
+#      - name: Set container image
+#        id: cann-image
+#        run: |
+#          image="ascendai/cann:${{ matrix.chip_type == '910b' &&  '8.5.0-910b-openeuler24.03-py3.11' || '8.5.0-310p-openeuler24.03-py3.11' }}"
+#          echo "image=${image}" >> "${GITHUB_OUTPUT}"
+#
+#      - name: Pull container image
+#        run: docker pull "${{ steps.cann-image.outputs.image }}"
+#
+#      - name: Build
+#        env:
+#          BUILD_TYPE: ${{ matrix.build }}
+#          SOC_TYPE: ascend${{ matrix.chip_type }}
+#          USE_ACL_GRAPH: ${{ matrix.use_acl_graph }}
+#        run: |
+#          HOST_UID=$(id -u)
+#          HOST_GID=$(id -g)
+#
+#          docker run --rm \
+#            -v "${PWD}:/workspace" \
+#            -w /workspace \
+#            -e SOC_TYPE=${SOC_TYPE} \
+#            -e BUILD_TYPE=${BUILD_TYPE} \
+#            -e USE_ACL_GRAPH=${USE_ACL_GRAPH} \
+#            "${{ steps.cann-image.outputs.image }}" \
+#            bash -lc '
+#              set -e
+#              yum install -y --setopt=install_weak_deps=False --setopt=tsflags=nodocs git gcc gcc-c++ make cmake openssl-devel
+#              yum clean all && rm -rf /var/cache/yum
+#              git config --global --add safe.directory "/workspace"
+#              export LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/$(uname -m)-linux/devlib/:${LD_LIBRARY_PATH}
+#              cmake -S . -B build \
+#                  -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
+#                  -DGGML_CANN=on \
+#                  -DSOC_TYPE=${SOC_TYPE} \
+#                  -DUSE_ACL_GRAPH=${USE_ACL_GRAPH}
+#              cmake --build build -j $(nproc)
+#
+#              chown -R '"${HOST_UID}"':'"${HOST_GID}"' /workspace/build
+#            '

.github/workflows/build-cpu.yml

@@ -0,0 +1,215 @@
+name: CI (cpu)
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/build-cpu.yml',
+      '.github/workflows/build-cmake-pkg.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+    ]
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/build-cpu.yml',
+      '.github/workflows/build-cmake-pkg.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp'
+    ]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
+
+jobs:
+  build-cmake-pkg:
+    uses: ./.github/workflows/build-cmake-pkg.yml
+
+  ubuntu:
+    strategy:
+      matrix:
+        include:
+          - build: 'x64'
+            os: ubuntu-22.04
+          - build: 'arm64'
+            os: ubuntu-24.04-arm
+
+    runs-on: ${{ matrix.os }}
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: cpu-${{ matrix.os }}
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Build Dependencies
+        id: build_depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y --no-install-recommends \
+            python3 python3-pip python3-dev python3-wheel \
+            libjpeg-dev build-essential libssl-dev \
+            git-lfs
+
+      - name: Toolchain workaround (GCC 14)
+        if: ${{ contains(matrix.os, 'ubuntu-24.04') }}
+        run: |
+          sudo apt-get install -y gcc-14 g++-14
+          echo "CC=gcc-14" >> "$GITHUB_ENV"
+          echo "CXX=g++-14" >> "$GITHUB_ENV"
+
+      - name: Python Dependencies
+        id: python_depends
+        run: |
+          export PIP_BREAK_SYSTEM_PACKAGES="1"
+          python3 -m pip install --upgrade pip setuptools
+          pip3 install ./gguf-py
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DGGML_RPC=ON
+          time cmake --build build --config Release -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
+      - name: Test llama2c conversion
+        id: llama2c_test
+        run: |
+          cd build
+          echo "Fetch tokenizer"
+          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/tok512.bin
+          echo "Fetch llama2c model"
+          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/stories260K.bin
+          ./bin/llama-convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
+          ./bin/llama-completion -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
+
+  windows:
+    runs-on: windows-2025
+
+    env:
+      OPENBLAS_VERSION: 0.3.23
+      SDE_VERSION: 9.33.0-2024-01-07
+      VULKAN_VERSION: 1.4.313.2
+
+    strategy:
+      matrix:
+        include:
+          - build: 'x64-cpu-static'
+            arch: 'x64'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF'
+          - build: 'x64-openblas'
+            arch: 'x64'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
+          - build: 'x64-vulkan'
+            arch: 'x64'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
+          - build: 'arm64'
+            arch: 'arm64'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: cpu-windows-2025-${{ matrix.build }}
+          variant: ccache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Download OpenBLAS
+        id: get_openblas
+        if: ${{ matrix.build == 'x64-openblas' }}
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/openblas.zip -L "https://github.com/xianyi/OpenBLAS/releases/download/v${env:OPENBLAS_VERSION}/OpenBLAS-${env:OPENBLAS_VERSION}-x64.zip"
+          curl.exe -o $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt -L "https://github.com/xianyi/OpenBLAS/raw/v${env:OPENBLAS_VERSION}/LICENSE"
+          mkdir $env:RUNNER_TEMP/openblas
+          tar.exe -xvf $env:RUNNER_TEMP/openblas.zip -C $env:RUNNER_TEMP/openblas
+          $vcdir = $(vswhere -latest -products * -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property installationPath)
+          $msvc = $(join-path $vcdir $('VC\Tools\MSVC\'+$(gc -raw $(join-path $vcdir 'VC\Auxiliary\Build\Microsoft.VCToolsVersion.default.txt')).Trim()))
+          $lib =  $(join-path $msvc 'bin\Hostx64\x64\lib.exe')
+          & $lib /machine:x64 "/def:${env:RUNNER_TEMP}/openblas/lib/libopenblas.def" "/out:${env:RUNNER_TEMP}/openblas/lib/openblas.lib" /name:openblas.dll
+
+      - name: Install Vulkan SDK
+        id: get_vulkan
+        if: ${{ matrix.build == 'x64-vulkan' }}
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
+          & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
+          Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
+          Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
+
+      - name: Install Ninja
+        id: install_ninja
+        run: |
+          choco install ninja
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -S . -B build ${{ matrix.defines }} `
+            -DLLAMA_BUILD_BORINGSSL=ON
+          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
+
+      - name: Add libopenblas.dll
+        id: add_libopenblas_dll
+        if: ${{ matrix.build == 'x64-openblas' }}
+        run: |
+          cp $env:RUNNER_TEMP/openblas/bin/libopenblas.dll ./build/bin/Release/openblas.dll
+          cp $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt ./build/bin/Release/OpenBLAS-${env:OPENBLAS_VERSION}.txt
+
+      - name: Test
+        id: cmake_test
+        if: ${{ matrix.arch == 'x64' }}
+        run: |
+          cd build
+          ctest -L main -C Release --verbose --timeout 900
+
+      # TODO: disabled for now, consider adding tests for all CPU variants instead
+      # - name: Test (Intel SDE)
+      #   id: cmake_test_sde
+      #   if: ${{ matrix.build == 'avx512-x64' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation
+      #   run: |
+      #     curl.exe -o $env:RUNNER_TEMP/sde.tar.xz -L "https://downloadmirror.intel.com/813591/sde-external-${env:SDE_VERSION}-win.tar.xz"
+      #     # for some weird reason windows tar doesn't like sde tar.xz
+      #     7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar.xz
+      #     7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar
+      #     $sde = $(join-path $env:RUNNER_TEMP sde-external-${env:SDE_VERSION}-win/sde.exe)
+      #     cd build
+      #     $env:LLAMA_SKIP_TESTS_SLOW_ON_EMULATOR = 1
+      #     & $sde -future -- ctest -L main -C Release --verbose --timeout 900

.github/workflows/build-cross.yml

@@ -287,7 +287,7 @@ jobs:
       #  id: cache-toolchain
       #  with:
       #    path: ./spacemit_toolchain
-      #    key: spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}-${{ runner.os }}
+      #    key: cache-gha-spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}-${{ runner.os }}
 
       - name: Setup SpacemiT Toolchain
         #if: steps.cache-toolchain.outputs.cache-hit != 'true'

.github/workflows/build-cuda-ubuntu.yml

@@ -0,0 +1,134 @@
+name: CI (CUDA, ubuntu)
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/build-cuda-ubuntu.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+      '**/*.cu',
+      '**/*.cuh'
+    ]
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/build-cuda-ubuntu.yml',
+      'ggml/src/ggml-cuda/**'
+    ]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
+
+jobs:
+  cuda:
+    runs-on: ubuntu-24.04
+    container: nvidia/cuda:12.6.2-devel-ubuntu24.04
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Install dependencies
+        env:
+          DEBIAN_FRONTEND: noninteractive
+        run: |
+          apt update
+          apt install -y cmake build-essential ninja-build libgomp1 git libssl-dev
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: cuda-ubuntu-24.04-cuda
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Build with CMake
+        # TODO: Remove GGML_CUDA_CUB_3DOT2 flag once CCCL 3.2 is bundled within CTK and that CTK version is used in this project
+        run: |
+          cmake -S . -B build -G Ninja \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DCMAKE_CUDA_ARCHITECTURES=89-real \
+            -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined \
+            -DGGML_NATIVE=OFF \
+            -DGGML_CUDA=ON \
+            -DGGML_CUDA_CUB_3DOT2=ON
+          cmake --build build
+
+  hip:
+    runs-on: ubuntu-22.04
+    container: rocm/dev-ubuntu-22.04:6.1.2
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libssl-dev rocwmma-dev
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: cuda-ubuntu-22.04-hip
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Build with native CMake HIP support
+        id: cmake_build
+        run: |
+          cmake -B build -S . \
+            -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \
+            -DGGML_HIP_ROCWMMA_FATTN=ON \
+            -DGPU_TARGETS="gfx1030" \
+            -DGGML_HIP=ON
+          cmake --build build --config Release -j $(nproc)
+
+  musa:
+    runs-on: ubuntu-22.04
+    container: mthreads/musa:rc4.3.0-devel-ubuntu22.04-amd64
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Dependencies
+        id: depends
+        run: |
+          apt-get update
+          apt-get install -y build-essential git cmake libssl-dev
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: cuda-ubuntu-22.04-musa
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Build with native CMake MUSA support
+        id: cmake_build
+        run: |
+          cmake -B build -S . \
+            -DGGML_MUSA=ON
+          time cmake --build build --config Release -j $(nproc)

.github/workflows/build-cuda-windows.yml

@@ -0,0 +1,162 @@
+name: CI (CUDA, windows)
+
+# TODO: this workflow is only triggered manually because it is very heavy on the CI
+#       when we provision dedicated windows runners, we can enable it for pushes too
+# note: running this workflow manually will populate the ccache for the release builds
+#       this can be used before merging a PR to speed up the release workflow
+on:
+  workflow_dispatch: # allows manual triggering
+
+# note: this will run in queue with the release workflow
+concurrency:
+  group: release
+  queue: max
+
+env:
+  GH_TOKEN: ${{ github.token }}
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
+
+jobs:
+  cuda:
+    runs-on: windows-2022
+
+    permissions:
+      actions: write
+
+    strategy:
+      matrix:
+        cuda: ['12.4', '13.3']
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: release-windows-2022-x64-cuda-${{ matrix.cuda }}
+
+      - name: Install Cuda Toolkit
+        uses: ./.github/actions/windows-setup-cuda
+        with:
+          cuda_version: ${{ matrix.cuda }}
+
+      - name: Install Ninja
+        id: install_ninja
+        run: |
+          choco install ninja
+
+      - name: Build
+        id: cmake_build
+        shell: cmd
+        # TODO: Remove GGML_CUDA_CUB_3DOT2 flag once CCCL 3.2 is bundled within CTK and that CTK version is used in this project
+        run: |
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
+          cmake -S . -B build -G "Ninja Multi-Config" ^
+            -DLLAMA_BUILD_SERVER=ON ^
+            -DLLAMA_BUILD_BORINGSSL=ON ^
+            -DGGML_NATIVE=OFF ^
+            -DGGML_BACKEND_DL=ON ^
+            -DGGML_CPU_ALL_VARIANTS=ON ^
+            -DGGML_CUDA=ON ^
+            -DGGML_RPC=ON ^
+            -DGGML_CUDA_CUB_3DOT2=ON
+          set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
+          cmake --build build --config Release -j %NINJA_JOBS% -t ggml
+          cmake --build build --config Release
+
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2022-x64-cuda-${{ matrix.cuda }}
+
+  hip:
+    runs-on: windows-2022
+
+    permissions:
+      actions: write
+
+    env:
+      # Make sure this is in sync with build-cache.yml
+      HIPSDK_INSTALLER_VERSION: "26.Q1"
+
+    strategy:
+      matrix:
+        include:
+          # sync with release.yml
+          - name: "radeon"
+            gpu_targets: "gfx1150;gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Grab rocWMMA package
+        id: grab_rocwmma
+        run: |
+          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/7.2.1/pool/main/r/rocwmma-dev/rocwmma-dev_2.2.0.70201-81~24.04_amd64.deb"
+          7z x rocwmma.deb
+          7z x data.tar
+
+      - name: Use ROCm Installation Cache
+        uses: actions/cache@v5
+        id: cache-rocm
+        with:
+          path: C:\Program Files\AMD\ROCm
+          key: cache-gha-rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
+
+      - name: Setup ROCm
+        if: steps.cache-rocm.outputs.cache-hit != 'true'
+        uses: ./.github/actions/windows-setup-rocm
+        with:
+          version: ${{ env.HIPSDK_INSTALLER_VERSION }}
+
+      - name: Verify ROCm
+        id: verify
+        run: |
+          # Find and test ROCm installation
+          $clangPath = Get-ChildItem 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | Select-Object -First 1
+          if (-not $clangPath) {
+            Write-Error "ROCm installation not found"
+            exit 1
+          }
+          & $clangPath.FullName --version
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          # TODO: this build does not match the build in release.yml, so we use a different cache key
+          #       ideally, the builds should match, similar to the CUDA build above so that we would be able
+          #       to populate the ccache for the release with manual runs of this workflow
+          #key: release-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
+          key: cuda-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
+
+      - name: Build
+        id: cmake_build
+        run: |
+          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
+          $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
+          cmake -G "Unix Makefiles" -B build -S . `
+            -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
+            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
+            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-7.2.1/include/" `
+            -DCMAKE_BUILD_TYPE=Release `
+            -DLLAMA_BUILD_BORINGSSL=ON `
+            -DROCM_DIR="${env:HIP_PATH}" `
+            -DGGML_HIP=ON `
+            -DGGML_HIP_ROCWMMA_FATTN=ON `
+            -DGPU_TARGETS="gfx1100"  `
+            -DGGML_RPC=ON
+          cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
+
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          #key: release-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}
+          key: cuda-windows-2022-x64-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}

.github/workflows/build-ibm.yml

@@ -0,0 +1,150 @@
+name: CI (ibm)
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/build-ibm.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp'
+    ]
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/build-ibm.yml',
+      'ggml/src/ggml-cpu/**'
+    ]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
+
+jobs:
+
+  ubuntu-24-s390x:
+    runs-on: ubuntu-24.04-s390x
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Build Dependencies
+        id: build_depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y --no-install-recommends \
+            python3 python3-pip python3-dev python3-wheel \
+            libjpeg-dev build-essential libssl-dev \
+            git-lfs
+
+      - name: Toolchain workaround (GCC 14)
+        run: |
+          sudo apt-get install -y gcc-14 g++-14
+          echo "CC=gcc-14" >> "$GITHUB_ENV"
+          echo "CXX=g++-14" >> "$GITHUB_ENV"
+
+      - name: Python Dependencies
+        id: python_depends
+        run: |
+          export PIP_BREAK_SYSTEM_PACKAGES="1"
+          python3 -m pip install --upgrade pip setuptools
+          pip3 install ./gguf-py
+
+      - name: Swap Endianness
+        id: endianness
+        run: |
+          for f in models/*.gguf; do
+            echo YES | python3 gguf-py/gguf/scripts/gguf_convert_endian.py $f big
+          done
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DGGML_RPC=ON
+          time cmake --build build --config Release -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
+      - name: Test llama2c (s390x)
+        id: llama2c_test_s390x
+        run: |
+          cd build
+          echo "Fetch llama2c big-endian model"
+          wget https://huggingface.co/ggml-org/models/resolve/main/tinyllamas/stories260K-be.gguf
+          ./bin/llama-completion -m stories260K-be.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
+
+  ubuntu-24-ppc64le:
+    runs-on: ubuntu-24.04-ppc64le
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Build Dependencies
+        id: build_depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y --no-install-recommends \
+            python3 python3-pip python3-dev python3-wheel \
+            libjpeg-dev build-essential libssl-dev \
+            git-lfs
+
+      - name: Toolchain workaround (GCC 14)
+        run: |
+          sudo apt-get install -y gcc-14 g++-14
+          echo "CC=gcc-14" >> "$GITHUB_ENV"
+          echo "CXX=g++-14" >> "$GITHUB_ENV"
+
+      - name: Python Dependencies
+        id: python_depends
+        run: |
+          export PIP_BREAK_SYSTEM_PACKAGES="1"
+          python3 -m pip install --upgrade pip setuptools
+          pip3 install ./gguf-py
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DGGML_RPC=ON
+          time cmake --build build --config Release -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
+      - name: Test llama2c conversion
+        id: llama2c_test
+        run: |
+          cd build
+          echo "Fetch tokenizer"
+          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/tok512.bin
+          echo "Fetch llama2c model"
+          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/stories260K.bin
+          ./bin/llama-convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
+          ./bin/llama-completion -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256

.github/workflows/build-msys.yml

@@ -15,9 +15,9 @@ concurrency:
 env:
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
-  LLAMA_LOG_COLORS: 1
-  LLAMA_LOG_PREFIX: 1
-  LLAMA_LOG_TIMESTAMPS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
 
 jobs:
   windows-msys2:
@@ -37,7 +37,7 @@ jobs:
       #- name: ccache
       #  uses: ggml-org/ccache-action@v1.2.16
       #  with:
-      #    key: windows-msys2
+      #    key: msys-windows-2025-x64
       #    variant: ccache
       #    evict-old-files: 1d
       #    save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}

.github/workflows/build-opencl.yml

@@ -0,0 +1,82 @@
+name: CI (opencl)
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/build-opencl.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+      '**/*.cl'
+    ]
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/build-opencl.yml',
+      'ggml/src/ggml-opencl/**'
+    ]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
+
+jobs:
+  windows-2025-opencl-adreno:
+    runs-on: windows-2025
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: opencl-windows-2025-x64
+          variant: ccache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Install Ninja
+        id: install_ninja
+        run: |
+          choco install ninja
+
+      - name: Install OpenCL Headers and Libs
+        id: install_opencl
+        run: |
+          git clone https://github.com/KhronosGroup/OpenCL-Headers
+          cd OpenCL-Headers
+          cmake -B build `
+            -DBUILD_TESTING=OFF `
+            -DOPENCL_HEADERS_BUILD_TESTING=OFF `
+            -DOPENCL_HEADERS_BUILD_CXX_TESTS=OFF `
+            -DCMAKE_INSTALL_PREFIX="$env:RUNNER_TEMP/opencl-arm64-release"
+          cmake --build build --target install
+          git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader
+          cd OpenCL-ICD-Loader
+          cmake -B build-arm64-release `
+            -A arm64 `
+            -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" `
+            -DCMAKE_INSTALL_PREFIX="$env:RUNNER_TEMP/opencl-arm64-release"
+          cmake --build build-arm64-release --target install --config release
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -S . -B build -G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON -DLLAMA_BUILD_BORINGSSL=ON
+          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}

.github/workflows/build-openvino.yml

@@ -29,9 +29,9 @@ concurrency:
 env:
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
-  LLAMA_LOG_COLORS: 1
-  LLAMA_LOG_PREFIX: 1
-  LLAMA_LOG_TIMESTAMPS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
 
 jobs:
   ubuntu-24-openvino:
@@ -67,7 +67,7 @@ jobs:
         if: runner.environment == 'github-hosted'
         uses: ggml-org/ccache-action@v1.2.21
         with:
-          key: ubuntu-24-openvino-${{ matrix.variant }}-no-preset-v1
+          key: openvino-ubuntu-24.04-${{ matrix.variant }}-no-preset-v1
           evict-old-files: 1d
           save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
@@ -84,7 +84,7 @@ jobs:
         id: cache-openvino
         with:
           path: ./openvino_toolkit
-          key: openvino-toolkit-v${{ env.OPENVINO_VERSION_FULL }}-${{ runner.os }}
+          key: cache-gha-openvino-toolkit-v${{ env.OPENVINO_VERSION_FULL }}-${{ runner.os }}
 
       - name: Setup OpenVINO Toolkit
         if: steps.cache-openvino.outputs.cache-hit != 'true'

.github/workflows/build-riscv.yml

@@ -29,11 +29,84 @@ concurrency:
 env:
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
-  LLAMA_LOG_COLORS: 1
-  LLAMA_LOG_PREFIX: 1
-  LLAMA_LOG_TIMESTAMPS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
 
 jobs:
+  ubuntu-cpu-riscv64-native:
+    runs-on: ubuntu-24.04-riscv
+
+    steps:
+      - name: Install dependencies
+        run: |
+          # Install necessary packages
+          sudo apt-get update
+          sudo apt-get install -y libssl-dev
+
+          # Set gcc-14 and g++-14 as the default compilers
+          sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-14 100
+          sudo update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-14 100
+
+          git lfs install
+
+      - name: Check environment
+        run: |
+          uname -a
+          gcc --version
+          g++ --version
+          ldd --version
+          cmake --version
+          rustc --version
+          env
+          echo "nproc=$(nproc)"
+
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      # note: sparing some ccache since these jobs run on dedicated runners that are not part of the organitzation
+      #- name: ccache
+      #  uses: ggml-org/ccache-action@afde29e5b5422e5da23cb1f639e8baecadeadfc3 # https://github.com/ggml-org/ccache-action/pull/1
+      #  with:
+      #    key: riscv-ubuntu-native
+      #    evict-old-files: 1d
+      #    save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_OPENMP=OFF \
+            -DLLAMA_BUILD_EXAMPLES=ON \
+            -DLLAMA_BUILD_TOOLS=ON \
+            -DLLAMA_BUILD_TESTS=ON \
+            -DCMAKE_C_COMPILER_LAUNCHER=ccache \
+            -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
+            -DGGML_RPC=ON \
+            -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
+            -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14
+
+          time cmake --build build --config Release -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
+      - name: Test llama2c conversion
+        id: llama2c_test
+        run: |
+          cd build
+          echo "Fetch tokenizer"
+          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/tok512.bin
+          echo "Fetch llama2c model"
+          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/stories260K.bin
+          ./bin/llama-convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
+          ./bin/llama-completion -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
+
   ubuntu-riscv64-native-sanitizer:
     runs-on: ubuntu-24.04-riscv
 
@@ -62,12 +135,13 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
-      - name: ccache
-        uses: ggml-org/ccache-action@afde29e5b5422e5da23cb1f639e8baecadeadfc3 # https://github.com/ggml-org/ccache-action/pull/1
-        with:
-          key: ubuntu-riscv64-native-sanitizer-${{ matrix.sanitizer }}-${{ matrix.build_type }}
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+      # note: sparing some ccache since these jobs run on dedicated runners that are not part of the organitzation
+      #- name: ccache
+      #  uses: ggml-org/ccache-action@afde29e5b5422e5da23cb1f639e8baecadeadfc3 # https://github.com/ggml-org/ccache-action/pull/1
+      #  with:
+      #    key: riscv-ubuntu-native-sanitizer-${{ matrix.sanitizer }}-${{ matrix.build_type }}
+      #    evict-old-files: 1d
+      #    save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Build
         id: cmake_build

.github/workflows/build-rpc.yml

@@ -0,0 +1,66 @@
+name: CI (rpc)
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/build-rpc.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp'
+    ]
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/build-rpc.yml',
+      'ggml/src/ggml-rpc/**'
+    ]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
+
+jobs:
+  ubuntu-latest-rpc:
+    runs-on: ubuntu-latest
+
+    continue-on-error: true
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential libssl-dev ninja-build
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build \
+            -G "Ninja" \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_RPC=ON
+          time cmake --build build --config Release -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose

.github/workflows/build-sanitize.yml

@@ -22,66 +22,65 @@ concurrency:
 env:
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
-  LLAMA_LOG_COLORS: 1
-  LLAMA_LOG_PREFIX: 1
-  LLAMA_LOG_TIMESTAMPS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
 
 jobs:
-  ubuntu-latest-sanitizer:
-    runs-on: ubuntu-latest
+  ctest:
+    runs-on: [self-hosted, X64, CPU, Linux]
 
     continue-on-error: true
 
     strategy:
       matrix:
         sanitizer: [ADDRESS, THREAD, UNDEFINED]
-        build_type: [Debug]
 
     steps:
       - name: Clone
         id: checkout
         uses: actions/checkout@v6
 
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: ubuntu-latest-sanitizer-${{ matrix.sanitizer }}
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      - name: Dependencies
-        id: depends
+      # with UNDEFINED sanitizer, we have to build in Debug to avoid GCC 13 false-positive warnings
+      - name: Build (undefined)
+        id: cmake_build_undefined
+        if: ${{ matrix.sanitizer == 'UNDEFINED' }}
         run: |
-          sudo apt-get update
-          sudo apt-get install build-essential libssl-dev
+          cmake -B build \
+            -DCMAKE_BUILD_TYPE=Debug \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
+            -DGGML_SANITIZE_${{ matrix.sanitizer }}=ON
+
+          cmake --build build --config Debug -j $(nproc)
 
       - name: Build
         id: cmake_build
-        if: ${{ matrix.sanitizer != 'THREAD' }}
+        if: ${{ matrix.sanitizer == 'ADDRESS' }}
         run: |
           cmake -B build \
-            -DLLAMA_FATAL_WARNINGS=ON \
+            -DCMAKE_BUILD_TYPE=RelWithDebInfo \
             -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
-            -DGGML_SANITIZE_${{ matrix.sanitizer }}=ON \
-            -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
+            -DGGML_SANITIZE_${{ matrix.sanitizer }}=ON
 
-          cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
+          cmake --build build --config RelWithDebInfo -j $(nproc)
 
       - name: Build (no OpenMP)
         id: cmake_build_no_openmp
         if: ${{ matrix.sanitizer == 'THREAD' }}
         run: |
           cmake -B build \
-            -DLLAMA_FATAL_WARNINGS=ON \
+            -DCMAKE_BUILD_TYPE=RelWithDebInfo \
             -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
             -DGGML_SANITIZE_${{ matrix.sanitizer }}=ON \
-            -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
             -DGGML_OPENMP=OFF
 
-          cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
+          cmake --build build --config RelWithDebInfo -j $(nproc)
 
       - name: Test
         id: cmake_test
+        # skip run in Debug - very slow
+        if: ${{ matrix.sanitizer != 'UNDEFINED' }}
         run: |
           cd build
-          ctest -L main --verbose --timeout 900
+          ctest -L main -E tokenizer --verbose --timeout 900

.github/workflows/build-self-hosted.yml

@@ -50,12 +50,12 @@ concurrency:
 env:
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
-  LLAMA_LOG_COLORS: 1
-  LLAMA_LOG_PREFIX: 1
-  LLAMA_LOG_TIMESTAMPS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
 
 jobs:
-  ggml-ci-nvidia-cuda:
+  gpu-cuda:
     runs-on: [self-hosted, Linux, NVIDIA]
 
     steps:
@@ -69,7 +69,7 @@ jobs:
           nvidia-smi
           GG_BUILD_CUDA=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
 
-  ggml-ci-nvidia-vulkan-cm:
+  gpu-vulkan-nvidia-cm:
     runs-on: [self-hosted, Linux, NVIDIA]
 
     steps:
@@ -83,7 +83,7 @@ jobs:
           vulkaninfo --summary
           GG_BUILD_VULKAN=1 GGML_VK_DISABLE_COOPMAT2=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
 
-  ggml-ci-nvidia-vulkan-cm2:
+  gpu-vulkan-nvidia-cm2:
     runs-on: [self-hosted, Linux, NVIDIA, COOPMAT2]
 
     steps:
@@ -97,7 +97,7 @@ jobs:
           vulkaninfo --summary
           GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
 
-  ggml-ci-nvidia-webgpu:
+  gpu-webgpu-nvidia:
     runs-on: [self-hosted, Linux, NVIDIA, X64]
 
     steps:
@@ -127,7 +127,7 @@ jobs:
             bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
 
   # TODO: provision AMX-compatible machine
-  #ggml-ci-cpu-amx:
+  #cpu-amx:
   #  runs-on: [self-hosted, Linux, CPU, AMX]
 
   #  steps:
@@ -141,7 +141,7 @@ jobs:
   #        bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
 
   # TODO: provision AMD GPU machine
-  # ggml-ci-amd-vulkan:
+  # amd-vulkan:
   #   runs-on: [self-hosted, Linux, AMD]
 
   #   steps:
@@ -156,7 +156,7 @@ jobs:
   #         GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
 
   # TODO: provision AMD GPU machine
-  # ggml-ci-amd-rocm:
+  # amd-rocm:
   #   runs-on: [self-hosted, Linux, AMD]
 
   #   steps:
@@ -170,7 +170,7 @@ jobs:
   #         amd-smi static
   #         GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS="gfx1101" bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
 
-  ggml-ci-mac-metal:
+  gpu-metal:
     runs-on: [self-hosted, macOS, ARM64]
 
     steps:
@@ -183,7 +183,7 @@ jobs:
         run: |
           GG_BUILD_METAL=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
 
-  ggml-ci-mac-webgpu:
+  gpu-webgpu-apple:
     runs-on: [self-hosted, macOS, ARM64]
 
     steps:
@@ -210,7 +210,7 @@ jobs:
           GG_BUILD_WEBGPU=1 GG_BUILD_WEBGPU_DAWN_PREFIX="$GITHUB_WORKSPACE/dawn" \
             bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
 
-  ggml-ci-mac-vulkan:
+  gpu-vulkan:
     runs-on: [self-hosted, macOS, ARM64]
 
     steps:
@@ -224,7 +224,7 @@ jobs:
           vulkaninfo --summary
           GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
 
-  ggml-ci-linux-intel-vulkan:
+  gpu-vulkan-intel-linux:
     runs-on: [self-hosted, Linux, Intel]
 
     steps:
@@ -240,7 +240,7 @@ jobs:
           vulkaninfo --summary
           GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
 
-  ggml-ci-win-intel-vulkan:
+  gpu-vulkan-intel-windows:
     runs-on: [self-hosted, Windows, X64, Intel]
 
     steps:
@@ -261,7 +261,7 @@ jobs:
           # a valid python environment for testing
           LLAMA_FATAL_WARNINGS=OFF GG_BUILD_NINJA=1 GG_BUILD_VULKAN=1 GG_BUILD_LOW_PERF=1 ./ci/run.sh ./results/llama.cpp ./mnt/llama.cpp
 
-  ggml-ci-intel-openvino-gpu-low-perf:
+  cpu-openvino-low-perf:
     runs-on: [self-hosted, Linux, Intel, OpenVINO]
 
     concurrency:
@@ -297,8 +297,8 @@ jobs:
           source ./openvino_toolkit/setupvars.sh
           GG_BUILD_OPENVINO=1 GGML_OPENVINO_DEVICE=GPU GG_BUILD_LOW_PERF=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
 
-  ggml-ci-arm64-cpu-low-perf:
-    runs-on: [self-hosted, Linux, ARM64, CPU]
+  cpu-any-low-perf:
+    runs-on: [self-hosted, CPU]
 
     steps:
       - name: Clone
@@ -310,8 +310,8 @@ jobs:
         run: |
           LLAMA_ARG_THREADS=$(nproc) GG_BUILD_LOW_PERF=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
 
-  ggml-ci-arm64-cpu-high-perf:
-    runs-on: [self-hosted, Linux, ARM64, CPU]
+  cpu-any-high-perf:
+    runs-on: [self-hosted, CPU]
 
     steps:
       - name: Clone
@@ -323,34 +323,82 @@ jobs:
         run: |
           LLAMA_ARG_THREADS=$(nproc) GG_BUILD_HIGH_PERF=1 GG_BUILD_NO_SVE=1 GG_BUILD_NO_BF16=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
 
-# TODO: not sure how to detect ARM flags on DGX Spark. currently get this error during cmake:
-#         CMake Warning at ggml/src/ggml-cpu/CMakeLists.txt:147 (message):
-#           ARM -march/-mcpu not found, -mcpu=native will be used
-#
-#       if we resolve this, we should be able to offload these jobs to the self-hosted runners
-#
-#  ggml-ci-arm64-cpu-high-perf-sve:
-#    runs-on: [self-hosted, Linux, ARM64, CPU]
-#
-#    steps:
-#      - name: Clone
-#        id: checkout
-#        uses: actions/checkout@v6
-#
-#      - name: Test
-#        id: ggml-ci
-#        run: |
-#          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_NO_BF16=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
-#
-#  ggml-ci-arm64-cpu-kleidiai:
-#    runs-on: [self-hosted, Linux, ARM64, CPU]
-#
-#    steps:
-#      - name: Clone
-#        id: checkout
-#        uses: actions/checkout@v6
-#
-#      - name: Test
-#        id: ggml-ci
-#        run: |
-#          GG_BUILD_KLEIDIAI=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+  cpu-arm64-graviton4:
+    runs-on: ah-ubuntu_22_04-c8g_8x
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Dependencies
+        id: depends
+        run: |
+          set -euxo pipefail
+          sudo apt-get update
+          sudo DEBIAN_FRONTEND=noninteractive NEEDRESTART_MODE=a \
+          apt-get install -y \
+          build-essential \
+          python3-venv \
+          gpg \
+          wget \
+          time \
+          git-lfs
+
+          git lfs install
+
+          # install the latest cmake
+          sudo install -d /usr/share/keyrings
+          wget -O - https://apt.kitware.com/keys/kitware-archive-latest.asc \
+            | gpg --dearmor \
+            | sudo tee /usr/share/keyrings/kitware-archive-keyring.gpg >/dev/null
+          echo 'deb [signed-by=/usr/share/keyrings/kitware-archive-keyring.gpg] https://apt.kitware.com/ubuntu/ jammy main' \
+            | sudo tee /etc/apt/sources.list.d/kitware.list
+          sudo apt-get update
+          sudo apt-get install -y cmake
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_NO_BF16=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
+  cpu-arm64-graviton4-kleidiai:
+    runs-on: ah-ubuntu_22_04-c8g_8x
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Dependencies
+        id: depends
+        run: |
+          set -euxo pipefail
+          sudo apt-get update
+          sudo DEBIAN_FRONTEND=noninteractive NEEDRESTART_MODE=a \
+          apt-get install -y \
+          build-essential \
+          python3-venv \
+          gpg \
+          wget \
+          time \
+          git-lfs
+
+          git lfs install
+
+          # install the latest cmake
+          sudo install -d /usr/share/keyrings
+          wget -O - https://apt.kitware.com/keys/kitware-archive-latest.asc \
+            | gpg --dearmor \
+            | sudo tee /usr/share/keyrings/kitware-archive-keyring.gpg >/dev/null
+          echo 'deb [signed-by=/usr/share/keyrings/kitware-archive-keyring.gpg] https://apt.kitware.com/ubuntu/ jammy main' \
+            | sudo tee /etc/apt/sources.list.d/kitware.list
+          sudo apt-get update
+          sudo apt-get install -y cmake
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          GG_BUILD_KLEIDIAI=1 \
+          GG_BUILD_EXTRA_TESTS_0=1 \
+          bash ./ci/run.sh ./tmp/results ./tmp/mnt

.github/workflows/build-sycl.yml

@@ -29,132 +29,134 @@ concurrency:
 env:
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
-  LLAMA_LOG_COLORS: 1
-  LLAMA_LOG_PREFIX: 1
-  LLAMA_LOG_TIMESTAMPS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
 
 jobs:
 
-  ubuntu-24-sycl:
-    strategy:
-      matrix:
-        build: [fp32, fp16]
-        include:
-          - build: fp32
-            fp16: OFF
-          - build: fp16
-            fp16: ON
+# TODO: this build is disabled to save Github Actions resources (https://github.com/ggml-org/llama.cpp/pull/23705)
+#       in order to enable it again, we have to provision dedicated runners  to run it
+#  ubuntu-24-sycl:
+#    strategy:
+#      matrix:
+#        build: [fp32]
+#        include:
+#          - build: fp32
+#            fp16: OFF
+#
+#    runs-on: ubuntu-24.04
+#
+#    env:
+#      ONEAPI_ROOT: /opt/intel/oneapi/
+#      ONEAPI_INSTALLER_VERSION: "2025.3.3"
+#      LEVEL_ZERO_VERSION: "1.28.2"
+#      LEVEL_ZERO_UBUNTU_VERSION: "u24.04"
+#
+#    continue-on-error: true
+#
+#    steps:
+#      - uses: actions/checkout@v6
+#
+#      - name: Use oneAPI Installation Cache
+#        uses: actions/cache@v5
+#        id: cache-sycl
+#        with:
+#          path: ${{ env.ONEAPI_ROOT }}
+#          key: cache-gha-oneAPI-${{ env.ONEAPI_INSTALLER_VERSION }}-${{ runner.os }}
+#
+#      - name: Download & Install oneAPI
+#        shell: bash
+#        if: steps.cache-sycl.outputs.cache-hit != 'true'
+#        run: |
+#          cd /tmp
+#          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/56f7923a-adb8-43f3-8b02-2b60fcac8cab/intel-deep-learning-essentials-2025.3.3.16_offline.sh -O intel-deep-learning-essentials_offline.sh
+#          sudo bash intel-deep-learning-essentials_offline.sh -s -a --silent --eula accept
+#
+#      - name: Install Level Zero SDK
+#        shell: bash
+#        run: |
+#          cd /tmp
+#          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero.deb
+#          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero-devel_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero-devel.deb
+#          sudo apt-get install -y ./level-zero.deb ./level-zero-devel.deb
+#
+#      - name: Clone
+#        id: checkout
+#        uses: actions/checkout@v6
+#
+#      - name: ccache
+#        uses: ggml-org/ccache-action@v1.2.21
+#        with:
+#          key: sycl-ubuntu-24-${{ matrix.build }}
+#          evict-old-files: 1d
+#          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+#
+#      - name: Build
+#        id: cmake_build
+#        run: |
+#          source /opt/intel/oneapi/setvars.sh
+#          cmake -B build \
+#            -G "Ninja" \
+#            -DCMAKE_BUILD_TYPE=Release \
+#            -DGGML_SYCL=ON \
+#            -DCMAKE_C_COMPILER=icx \
+#            -DCMAKE_CXX_COMPILER=icpx \
+#            -DLLAMA_OPENSSL=OFF \
+#            -DGGML_NATIVE=OFF \
+#            -DGGML_SYCL_F16=${{ matrix.fp16 }}
+#          time cmake --build build --config Release -j $(nproc)
 
-    runs-on: ubuntu-24.04
-
-    env:
-      ONEAPI_ROOT: /opt/intel/oneapi/
-      ONEAPI_INSTALLER_VERSION: "2025.3.3"
-      LEVEL_ZERO_VERSION: "1.28.2"
-      LEVEL_ZERO_UBUNTU_VERSION: "u24.04"
-
-    continue-on-error: true
-
-    steps:
-      - uses: actions/checkout@v6
-
-      - name: Use oneAPI Installation Cache
-        uses: actions/cache@v5
-        id: cache-sycl
-        with:
-          path: ${{ env.ONEAPI_ROOT }}
-          key: oneAPI-${{ env.ONEAPI_INSTALLER_VERSION }}-${{ runner.os }}
-
-      - name: Download & Install oneAPI
-        shell: bash
-        if: steps.cache-sycl.outputs.cache-hit != 'true'
-        run: |
-          cd /tmp
-          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/56f7923a-adb8-43f3-8b02-2b60fcac8cab/intel-deep-learning-essentials-2025.3.3.16_offline.sh -O intel-deep-learning-essentials_offline.sh
-          sudo bash intel-deep-learning-essentials_offline.sh -s -a --silent --eula accept
-
-      - name: Install Level Zero SDK
-        shell: bash
-        run: |
-          cd /tmp
-          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero.deb
-          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero-devel_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero-devel.deb
-          sudo apt-get install -y ./level-zero.deb ./level-zero-devel.deb
-
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: ubuntu-24-sycl-${{ matrix.build }}
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      - name: Build
-        id: cmake_build
-        run: |
-          source /opt/intel/oneapi/setvars.sh
-          cmake -B build \
-            -G "Ninja" \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DGGML_SYCL=ON \
-            -DCMAKE_C_COMPILER=icx \
-            -DCMAKE_CXX_COMPILER=icpx \
-            -DLLAMA_OPENSSL=OFF \
-            -DGGML_NATIVE=OFF \
-            -DGGML_SYCL_F16=${{ matrix.fp16 }}
-          time cmake --build build --config Release -j $(nproc)
-
-  windows-latest-sycl:
-    runs-on: windows-2022
-
-    defaults:
-      run:
-        shell: bash
-
-    env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b60765d1-2b85-4e85-86b6-cb0e9563a699/intel-deep-learning-essentials-2025.3.3.18_offline.exe
-      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
-      LEVEL_ZERO_SDK_URL: https://github.com/oneapi-src/level-zero/releases/download/v1.28.2/level-zero-win-sdk-1.28.2.zip
-      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
-      ONEAPI_INSTALLER_VERSION: "2025.3.3"
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Use oneAPI Installation Cache
-        uses: actions/cache@v5
-        id: cache-sycl
-        with:
-          path: ${{ env.ONEAPI_ROOT }}
-          key: oneAPI-${{ env.ONEAPI_INSTALLER_VERSION }}-${{ runner.os }}
-
-      - name: Download & Install oneAPI
-        shell: bash
-        if: steps.cache-sycl.outputs.cache-hit != 'true'
-        run: |
-          scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
-
-      - name: Install Level Zero SDK
-        shell: pwsh
-        run: |
-          Invoke-WebRequest -Uri "${{ env.LEVEL_ZERO_SDK_URL }}" -OutFile "level-zero-win-sdk.zip"
-          Expand-Archive -Path "level-zero-win-sdk.zip" -DestinationPath "C:/level-zero-sdk" -Force
-          "LEVEL_ZERO_V1_SDK_PATH=C:/level-zero-sdk" | Out-File -FilePath $env:GITHUB_ENV -Append
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: windows-latest-sycl
-          variant: ccache
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      # TODO: add ssl support ; we will also need to modify win-build-sycl.bat to accept user-specified args
-
-      - name: Build
-        id: cmake_build
-        run:  examples/sycl/win-build-sycl.bat
+# TODO: this build is disabled to save Github Actions resources (https://github.com/ggml-org/llama.cpp/pull/23705)
+#       in order to enable it again, we have to provision dedicated runners  to run it
+#  windows-latest-sycl:
+#    runs-on: windows-2022
+#
+#    defaults:
+#      run:
+#        shell: bash
+#
+#    env:
+#      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b60765d1-2b85-4e85-86b6-cb0e9563a699/intel-deep-learning-essentials-2025.3.3.18_offline.exe
+#      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
+#      LEVEL_ZERO_SDK_URL: https://github.com/oneapi-src/level-zero/releases/download/v1.28.2/level-zero-win-sdk-1.28.2.zip
+#      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
+#      ONEAPI_INSTALLER_VERSION: "2025.3.3"
+#    steps:
+#      - name: Clone
+#        id: checkout
+#        uses: actions/checkout@v6
+#
+#      - name: Use oneAPI Installation Cache
+#        uses: actions/cache@v5
+#        id: cache-sycl
+#        with:
+#          path: ${{ env.ONEAPI_ROOT }}
+#          key: cache-gha-oneAPI-${{ env.ONEAPI_INSTALLER_VERSION }}-${{ runner.os }}
+#
+#      - name: Download & Install oneAPI
+#        shell: bash
+#        if: steps.cache-sycl.outputs.cache-hit != 'true'
+#        run: |
+#          scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
+#
+#      - name: Install Level Zero SDK
+#        shell: pwsh
+#        run: |
+#          Invoke-WebRequest -Uri "${{ env.LEVEL_ZERO_SDK_URL }}" -OutFile "level-zero-win-sdk.zip"
+#          Expand-Archive -Path "level-zero-win-sdk.zip" -DestinationPath "C:/level-zero-sdk" -Force
+#          "LEVEL_ZERO_V1_SDK_PATH=C:/level-zero-sdk" | Out-File -FilePath $env:GITHUB_ENV -Append
+#
+#      - name: ccache
+#        uses: ggml-org/ccache-action@v1.2.21
+#        with:
+#          key: sycl-windows-latest
+#          variant: ccache
+#          evict-old-files: 1d
+#          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+#
+#      # TODO: add ssl support ; we will also need to modify win-build-sycl.bat to accept user-specified args
+#
+#      - name: Build
+#        id: cmake_build
+#        run:  examples/sycl/win-build-sycl.bat

.github/workflows/build-vulkan.yml

@@ -31,26 +31,64 @@ concurrency:
 env:
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
-  LLAMA_LOG_COLORS: 1
-  LLAMA_LOG_PREFIX: 1
-  LLAMA_LOG_TIMESTAMPS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
 
 jobs:
-  ubuntu-24-vulkan-llvmpipe:
-    runs-on: ubuntu-24.04
+  ubuntu:
+    strategy:
+      matrix:
+        include:
+          - build: 'x64'
+            os: ubuntu-24.04
+          - build: 'arm64'
+            os: ubuntu-24.04-arm
+
+    runs-on: ${{ matrix.os }}
 
     steps:
       - name: Clone
         id: checkout
         uses: actions/checkout@v6
 
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y gcc-14 g++-14 build-essential glslc libvulkan-dev spirv-headers libssl-dev ninja-build
+          echo "CC=gcc-14" >> "$GITHUB_ENV"
+          echo "CXX=g++-14" >> "$GITHUB_ENV"
+
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
         with:
-          key: ubuntu-24-vulkan-llvmpipe
+          key: vulkan-${{ matrix.os }}-new
+          variant: ccache
           evict-old-files: 1d
           save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
+      - name: Configure
+        id: cmake_configure
+        run: |
+          cmake -B build \
+            -G "Ninja" \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_VULKAN=ON
+
+      - name: Build
+        id: cmake_build
+        run: |
+          time cmake --build build -j $(nproc)
+
+  ubuntu-llvmpipe:
+    runs-on: ubuntu-24.04
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
       - name: Dependencies
         id: depends
         run: |
@@ -68,7 +106,7 @@ jobs:
         id: cache-sdk
         with:
           path: ./vulkan_sdk
-          key: vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}
+          key: cache-gha-vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}
 
       - name: Setup Vulkan SDK
         if: steps.cache-sdk.outputs.cache-hit != 'true'
@@ -77,6 +115,13 @@ jobs:
           path: ./vulkan_sdk
           version: ${{ env.VULKAN_SDK_VERSION }}
 
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: vulkan-ubuntu-24.04-llvmpipe
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
       - name: Build
         id: cmake_build
         run: |

.github/workflows/build-webgpu.yml

@@ -0,0 +1,181 @@
+name: CI (webgpu)
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/build-webgpu.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+      '**/*.wgsl'
+    ]
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/build-webgpu.yml',
+      'ggml/src/ggml-webgpu/**'
+    ]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
+
+jobs:
+  macos:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: webgpu-macos-latest
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Dawn Dependency
+        id: dawn-depends
+        run: |
+          DAWN_VERSION="v20260317.182325"
+          DAWN_OWNER="google"
+          DAWN_REPO="dawn"
+          DAWN_ASSET_NAME="Dawn-18eb229ef5f707c1464cc581252e7603c73a3ef0-macos-latest-Release"
+          echo "Fetching release asset from https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
+          curl -L -o artifact.tar.gz \
+            "https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
+          mkdir dawn
+          tar -xvf artifact.tar.gz -C dawn --strip-components=1
+
+      - name: Build
+        id: cmake_build
+        run: |
+          export CMAKE_PREFIX_PATH=dawn
+          cmake -B build -G "Ninja" -DCMAKE_BUILD_TYPE=Release -DGGML_WEBGPU=ON -DGGML_METAL=OFF -DGGML_BLAS=OFF
+          time cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
+  ubuntu:
+    runs-on: ubuntu-24.04
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: webgpu-ubuntu-24.04
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo add-apt-repository -y ppa:kisak/kisak-mesa
+          sudo apt-get update -y
+          sudo apt-get install -y build-essential mesa-vulkan-drivers \
+            libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libssl-dev
+
+      - name: Dawn Dependency
+        id: dawn-depends
+        run: |
+          sudo apt-get install -y libxrandr-dev libxinerama-dev libxcursor-dev mesa-common-dev libx11-xcb-dev libxi-dev
+          DAWN_VERSION="v20260317.182325"
+          DAWN_OWNER="google"
+          DAWN_REPO="dawn"
+          DAWN_ASSET_NAME="Dawn-18eb229ef5f707c1464cc581252e7603c73a3ef0-ubuntu-latest-Release"
+          echo "Fetching release asset from https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
+          curl -L -o artifact.tar.gz \
+            "https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
+          mkdir dawn
+          tar -xvf artifact.tar.gz -C dawn --strip-components=1
+
+      - name: Build
+        id: cmake_build
+        run: |
+          export Dawn_DIR=dawn/lib64/cmake/Dawn
+          cmake -B build \
+            -DGGML_WEBGPU=ON
+          time cmake --build build --config Release -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          # This is using llvmpipe and runs slower than other backends
+          # test-backend-ops is too slow on llvmpipe, skip it
+          ctest -L main -E test-backend-ops --verbose --timeout 900
+
+  ubuntu-wasm:
+    strategy:
+      matrix:
+        include:
+          - build: 'x64'
+            os: ubuntu-24.04
+          - build: 'arm64'
+            os: ubuntu-24.04-arm
+
+    runs-on: ${{ matrix.os }}
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: webgpu-${{ matrix.os }}-wasm
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Install Emscripten
+        run: |
+          git clone https://github.com/emscripten-core/emsdk.git
+          cd emsdk
+          ./emsdk install latest
+          ./emsdk activate latest
+
+      - name: Fetch emdawnwebgpu
+        run: |
+          DAWN_TAG="v20260317.182325"
+          EMDAWN_PKG="emdawnwebgpu_pkg-${DAWN_TAG}.zip"
+          echo "Downloading ${EMDAWN_PKG}"
+          curl -L -o emdawn.zip \
+            "https://github.com/google/dawn/releases/download/${DAWN_TAG}/${EMDAWN_PKG}"
+          unzip emdawn.zip
+
+      - name: Build WASM WebGPU
+        run: |
+          source emsdk/emsdk_env.sh
+          emcmake cmake -B build-wasm \
+            -G "Ninja" \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_WEBGPU=ON \
+            -DLLAMA_OPENSSL=OFF \
+            -DEMDAWNWEBGPU_DIR=emdawnwebgpu_pkg
+
+          time cmake --build build-wasm --config Release --target test-backend-ops -j $(nproc)

.github/workflows/hip-quality-check.yml

@@ -28,9 +28,9 @@ concurrency:
 env:
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
-  LLAMA_LOG_COLORS: 1
-  LLAMA_LOG_PREFIX: 1
-  LLAMA_LOG_TIMESTAMPS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
 
 jobs:
   ubuntu-22-hip-quality-check:
@@ -50,7 +50,7 @@ jobs:
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
         with:
-          key: ubuntu-22-hip-quality-check
+          key: hip-quality-check-ubuntu-22.04
           evict-old-files: 1d
           save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 

.github/workflows/server-sanitize.yml

@@ -26,10 +26,10 @@ on:
     ]
 
 env:
-  LLAMA_LOG_COLORS: 1
-  LLAMA_LOG_PREFIX: 1
-  LLAMA_LOG_TIMESTAMPS: 1
-  LLAMA_LOG_VERBOSITY: 10
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
+  LLAMA_ARG_LOG_VERBOSITY: 10
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
@@ -37,7 +37,7 @@ concurrency:
 
 jobs:
   server:
-    runs-on: ubuntu-latest
+    runs-on: [self-hosted, CPU, Linux, llama-server]
 
     strategy:
       matrix:
@@ -46,19 +46,19 @@ jobs:
       fail-fast: false
 
     steps:
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get -y install \
-            build-essential \
-            xxd \
-            git \
-            cmake \
-            curl \
-            wget \
-            language-pack-en \
-            libssl-dev
+      #- name: Dependencies
+      #  id: depends
+      #  run: |
+      #    sudo apt-get update
+      #    sudo apt-get -y install \
+      #      build-essential \
+      #      xxd \
+      #      git \
+      #      cmake \
+      #      curl \
+      #      wget \
+      #      language-pack-en \
+      #      libssl-dev
 
       - name: Clone
         id: checkout

.github/workflows/server-self-hosted.yml

@@ -29,10 +29,10 @@ on:
     ]
 
 env:
-  LLAMA_LOG_COLORS: 1
-  LLAMA_LOG_PREFIX: 1
-  LLAMA_LOG_TIMESTAMPS: 1
-  LLAMA_LOG_VERBOSITY: 10
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
+  LLAMA_ARG_LOG_VERBOSITY: 10
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}

.github/workflows/server.yml

@@ -44,25 +44,20 @@ on:
     ]
 
 env:
-  LLAMA_LOG_COLORS: 1
-  LLAMA_LOG_PREFIX: 1
-  LLAMA_LOG_TIMESTAMPS: 1
-  LLAMA_LOG_VERBOSITY: 10
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
+  LLAMA_ARG_LOG_VERBOSITY: 10
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
   cancel-in-progress: true
 
 jobs:
-  ui-build:
-    name: Build Web UI
-    uses: ./.github/workflows/ui-build.yml
+  ubuntu:
+    runs-on: ubuntu-24.04
 
-  server:
-    runs-on: ubuntu-latest
-    needs: ui-build
-
-    name: server (${{ matrix.wf_name }})
+    name: ubuntu (${{ matrix.wf_name }})
     strategy:
       matrix:
         build_type: [Release]
@@ -98,17 +93,17 @@ jobs:
           fetch-depth: 0
           ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
 
-      - name: Download built UI
-        uses: actions/download-artifact@v7
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
         with:
-          name: ui-build
-          path: tools/ui/dist
+          key: server-ubuntu-24.04-x64
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Build
         id: cmake_build
         run: |
           cmake -B build \
-            -DLLAMA_BUILD_BORINGSSL=ON \
             -DGGML_SCHED_NO_REALLOC=ON
           cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
 
@@ -135,8 +130,8 @@ jobs:
           export ${{ matrix.extra_args }}
           SLOW_TESTS=1 pytest -v -x
 
-  server-windows:
-    runs-on: windows-2022
+  windows:
+    runs-on: windows-2025
 
     steps:
       - name: Clone
@@ -146,16 +141,24 @@ jobs:
           fetch-depth: 0
           ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
 
-      - name: Setup Node.js
-        uses: actions/setup-node@v6
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
         with:
-          node-version: "24"
+          key: server-windows-2025-x64
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Build
         id: cmake_build
+        shell: cmd
         run: |
-          cmake -B build -DLLAMA_BUILD_BORINGSSL=ON -DGGML_SCHED_NO_REALLOC=ON
-          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS} --target llama-server
+          cmake -B build -G "Ninja Multi-Config" ^
+            -DCMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake ^
+            -DCMAKE_BUILD_TYPE=Release ^
+            -DLLAMA_BUILD_BORINGSSL=ON ^
+            -DGGML_SCHED_NO_REALLOC=ON
+          set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
+          cmake --build build --config Release -j %NINJA_JOBS% --target llama-server
 
       - name: Python setup
         id: setup_python

.github/workflows/ui-build-self-hosted.yml

@@ -0,0 +1,43 @@
+name: UI Build (self-hosted)
+
+on:
+  workflow_call:
+
+jobs:
+  build:
+    runs-on: [self-hosted, fast]
+    env:
+      BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v6
+
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
+        with:
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
+
+      - name: Install dependencies
+        run: npm ci
+        working-directory: tools/ui
+
+      - name: Build application
+        run: npm run build
+        working-directory: tools/ui
+
+      - name: Generate checksums
+        run: |
+          cd tools/ui/dist
+          for f in *; do
+            sha256sum "$f" | awk '{print $1, $2}' >> checksums.txt
+          done
+
+      - name: Upload built UI
+        uses: actions/upload-artifact@v6
+        with:
+          name: ui-build
+          path: tools/ui/dist/
+          retention-days: 1

.github/workflows/ui-build.yml

@@ -5,7 +5,7 @@ on:
 
 jobs:
   build:
-    runs-on: [self-hosted, fast]
+    runs-on: ubuntu-slim
     env:
       BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
 

.github/workflows/ui-publish.yml

@@ -20,7 +20,7 @@ jobs:
   publish:
     name: Publish UI Static Output
     needs: build
-    runs-on: ubuntu-24.04-arm
+    runs-on: ubuntu-slim
 
     permissions:
       contents: read

.github/workflows/ui-self-hosted.yml

@@ -16,7 +16,7 @@ on:
       - master
     paths: [
       '.github/workflows/ui-self-hosted.yml',
-      '.github/workflows/ui-build.yml',
+      '.github/workflows/ui-build-self-hosted.yml',
       'tools/ui/**.*',
       'tools/server/tests/**.*'
     ]
@@ -24,16 +24,16 @@ on:
     types: [opened, synchronize, reopened]
     paths: [
       '.github/workflows/ui-self-hosted.yml',
-      '.github/workflows/ui-build.yml',
+      '.github/workflows/ui-build-self-hosted.yml',
       'tools/ui/**.*',
       'tools/server/tests/**.*'
     ]
 
 env:
-  LLAMA_LOG_COLORS: 1
-  LLAMA_LOG_PREFIX: 1
-  LLAMA_LOG_TIMESTAMPS: 1
-  LLAMA_LOG_VERBOSITY: 10
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
+  LLAMA_ARG_LOG_VERBOSITY: 10
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
@@ -42,7 +42,7 @@ concurrency:
 jobs:
   ui-build:
     name: Build static output
-    uses: ./.github/workflows/ui-build.yml
+    uses: ./.github/workflows/ui-build-self-hosted.yml
 
   ui-checks:
     name: Checks

.github/workflows/ui.yml

@@ -26,10 +26,10 @@ on:
     ]
 
 env:
-  LLAMA_LOG_COLORS: 1
-  LLAMA_LOG_PREFIX: 1
-  LLAMA_LOG_TIMESTAMPS: 1
-  LLAMA_LOG_VERBOSITY: 10
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
+  LLAMA_ARG_LOG_VERBOSITY: 10
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}

CMakeLists.txt

@@ -222,19 +222,6 @@ if (LLAMA_BUILD_APP)
     add_subdirectory(app)
 endif()
 
-# Automatically add all files from the 'licenses' directory
-file(GLOB EXTRA_LICENSES "${CMAKE_SOURCE_DIR}/licenses/LICENSE-*")
-
-foreach(FILE_PATH ${EXTRA_LICENSES})
-    get_filename_component(FILE_NAME "${FILE_PATH}" NAME)
-    string(REGEX REPLACE "^LICENSE-" "" NAME "${FILE_NAME}")
-    license_add_file("${NAME}" "${FILE_PATH}")
-endforeach()
-
-if (LLAMA_BUILD_COMMON)
-    license_generate(llama-common)
-endif()
-
 #
 # install
 #

CONTRIBUTING.md

@@ -63,6 +63,7 @@ After submitting your PR:
 - Optionally pick a `<module>` from here: https://github.com/ggml-org/llama.cpp/wiki/Modules
 - Let other maintainers merge their own PRs
 - When merging a PR, make sure you have a good understanding of the changes
+- If a PR does not warrant a new release, add `[no release]` in the squashed commit to spare CI resources
 - Be mindful of maintenance: most of the work going into a feature happens after the PR is merged. If the PR author is not committed to contribute long-term, someone else needs to take responsibility (you)
 
 Maintainers reserve the right to decline review or close pull requests for any reason, without any questions, particularly under any of the following conditions:

app/CMakeLists.txt

@@ -15,6 +15,17 @@ target_link_libraries(${TARGET} PRIVATE
 )
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
+# Automatically add all files from the 'licenses' directory
+file(GLOB EXTRA_LICENSES "${CMAKE_SOURCE_DIR}/licenses/LICENSE-*")
+
+foreach(FILE_PATH ${EXTRA_LICENSES})
+    get_filename_component(FILE_NAME "${FILE_PATH}" NAME)
+    string(REGEX REPLACE "^LICENSE-" "" NAME "${FILE_NAME}")
+    license_add_file("${NAME}" "${FILE_PATH}")
+endforeach()
+
+license_generate(${TARGET})
+
 if(LLAMA_TOOLS_INSTALL)
     install(TARGETS ${TARGET} RUNTIME)
 endif()

app/llama.cpp

@@ -5,6 +5,9 @@
 #include <string>
 #include <vector>
 
+// embedded data generated by cmake
+extern const char * LICENSES[];
+
 // visible
 int llama_server(int argc, char ** argv);
 int llama_cli(int argc, char ** argv);
@@ -17,8 +20,23 @@ int llama_fit_params(int argc, char ** argv);
 int llama_quantize(int argc, char ** argv);
 int llama_perplexity(int argc, char ** argv);
 
+// hands the update over to the install script, which downloads and swaps the binary
+static int llama_update(int argc, char ** argv) {
+    (void) argc;
+    (void) argv;
+
+#if defined(_WIN32)
+    return system("powershell -NoProfile -ExecutionPolicy Bypass -Command \"irm https://llama.app/install.ps1 | iex\"");
+#else
+    return system("curl -fsSL https://llama.app/install.sh | sh");
+#endif
+}
+
+static const char * progname;
+
 static int help(int argc, char ** argv);
 static int version(int argc, char ** argv);
+static int licenses(int argc, char ** argv);
 
 struct command {
     const char * name;
@@ -31,14 +49,16 @@ struct command {
 static const command cmds[] = {
     {"serve",         "HTTP API server",                                    {"server"},   false, llama_server       },
     {"cli",           "Command-line interactive interface",                 {"client"},   false, llama_cli          },
+    {"update",        "Update llama to the latest release",                 {},           false, llama_update       },
     {"completion",    "Text completion",                                    {"complete"}, true,  llama_completion   },
     {"bench",         "Benchmark prompt processing and text generation",    {},           true,  llama_bench        },
     {"batched-bench", "Benchmark batched decoding performance",             {},           true,  llama_batched_bench},
     {"fit-params",    "Compute parameters to fit a model in device memory", {},           true,  llama_fit_params   },
     {"quantize",      "Quantize a model",                                   {},           true,  llama_quantize     },
     {"perplexity",    "Compute model perplexity and KL divergence",         {},           true,  llama_perplexity   },
-    {"version",       "Show version",                                       {},           true,  version            },
-    {"help",          "Show available commands",                            {},           true,  help               },
+    {"version",       "Show version",                                       {},           false, version            },
+    {"licenses",      "Show third-party licenses",                          {"credits"},  false, licenses           },
+    {"help",          "Show available commands",                            {},           false, help               },
 };
 
 static int version(int argc, char ** argv) {
@@ -46,17 +66,29 @@ static int version(int argc, char ** argv) {
     return 0;
 }
 
+static int licenses(int argc, char ** argv) {
+    for (int i = 0; LICENSES[i]; ++i) {
+        printf("%s\n", LICENSES[i]);
+    }
+    return 0;
+}
+
 static int help(int argc, char ** argv) {
     const bool show_all = argc >= 2 && std::string(argv[1]) == "all";
 
-    printf("Usage: llama <command> [options]\n\nAvailable commands:\n");
+    printf("Usage: %s <command> [options]\n\nAvailable commands:\n", progname);
 
     for (const auto & cmd : cmds) {
         if (show_all || !cmd.hidden) {
             printf("  %-15s %s\n", cmd.name, cmd.desc);
         }
     }
-    printf("\nRun 'llama <command> --help' for command-specific usage.\n");
+    printf("\n");
+
+    if (!show_all) {
+        printf("Run '%s help all' to show additional commands.\n", progname);
+    }
+    printf("Run '%s <command> --help' for command-specific usage.\n", progname);
 
     return 0;
 }
@@ -74,13 +106,13 @@ static bool matches(const std::string & arg, const command & cmd) {
 }
 
 int main(int argc, char ** argv) {
+    progname = argv[0];
+
     const std::string arg = argc >= 2 ? argv[1] : "help";
 
     for (const auto & cmd : cmds) {
         if (matches(arg, cmd)) {
-
-            // router spawns children through this same binary, it needs the
-            // subcommand to relaunch as 'llama serve' and not bare options
+            // keep cmd.name so the router's child processes re-invoke correctly
 #ifdef _WIN32
             _putenv_s("LLAMA_APP_CMD", cmd.name);
 #else

build-xcframework.sh

@@ -8,6 +8,7 @@ TVOS_MIN_OS_VERSION=16.4
 
 BUILD_SHARED_LIBS=OFF
 LLAMA_BUILD_APP=OFF
+LLAMA_BUILD_COMMON=OFF
 LLAMA_BUILD_EXAMPLES=OFF
 LLAMA_BUILD_TOOLS=OFF
 LLAMA_BUILD_TESTS=OFF
@@ -33,6 +34,7 @@ COMMON_CMAKE_ARGS=(
     -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
     -DBUILD_SHARED_LIBS=${BUILD_SHARED_LIBS}
     -DLLAMA_BUILD_APP=${LLAMA_BUILD_APP}
+    -DLLAMA_BUILD_COMMON=${LLAMA_BUILD_COMMON}
     -DLLAMA_BUILD_EXAMPLES=${LLAMA_BUILD_EXAMPLES}
     -DLLAMA_BUILD_TOOLS=${LLAMA_BUILD_TOOLS}
     -DLLAMA_BUILD_TESTS=${LLAMA_BUILD_TESTS}
@@ -416,7 +418,7 @@ cmake -B build-ios-sim -G Xcode \
     -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
     -DLLAMA_OPENSSL=OFF \
     -S .
-cmake --build build-ios-sim --config Release -- -quiet
+cmake --build build-ios-sim --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
 
 echo "Building for iOS devices..."
 cmake -B build-ios-device -G Xcode \
@@ -430,7 +432,7 @@ cmake -B build-ios-device -G Xcode \
     -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
     -DLLAMA_OPENSSL=OFF \
     -S .
-cmake --build build-ios-device --config Release -- -quiet
+cmake --build build-ios-device --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
 
 echo "Building for macOS..."
 cmake -B build-macos -G Xcode \
@@ -441,7 +443,7 @@ cmake -B build-macos -G Xcode \
     -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
     -DLLAMA_OPENSSL=OFF \
     -S .
-cmake --build build-macos --config Release -- -quiet
+cmake --build build-macos --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
 
 echo "Building for visionOS..."
 cmake -B build-visionos -G Xcode \
@@ -456,7 +458,7 @@ cmake -B build-visionos -G Xcode \
     -DLLAMA_OPENSSL=OFF \
     -DLLAMA_BUILD_SERVER=OFF \
     -S .
-cmake --build build-visionos --config Release -- -quiet
+cmake --build build-visionos --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
 
 echo "Building for visionOS simulator..."
 cmake -B build-visionos-sim -G Xcode \
@@ -471,7 +473,7 @@ cmake -B build-visionos-sim -G Xcode \
     -DLLAMA_OPENSSL=OFF \
     -DLLAMA_BUILD_SERVER=OFF \
     -S .
-cmake --build build-visionos-sim --config Release -- -quiet
+cmake --build build-visionos-sim --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
 
 # Add tvOS builds (might need the same u_int definitions as watchOS and visionOS)
 echo "Building for tvOS simulator..."
@@ -487,7 +489,7 @@ cmake -B build-tvos-sim -G Xcode \
     -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
     -DLLAMA_OPENSSL=OFF \
     -S .
-cmake --build build-tvos-sim --config Release -- -quiet
+cmake --build build-tvos-sim --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
 
 echo "Building for tvOS devices..."
 cmake -B build-tvos-device -G Xcode \
@@ -502,7 +504,7 @@ cmake -B build-tvos-device -G Xcode \
     -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
     -DLLAMA_OPENSSL=OFF \
     -S .
-cmake --build build-tvos-device --config Release -- -quiet
+cmake --build build-tvos-device --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet
 
 # Setup frameworks and copy binaries and headers
 echo "Setting up framework structures..."

ci/run.sh

@@ -66,6 +66,8 @@ fi
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON"
+else
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=OFF"
 fi
 
 if [ ! -z ${GG_BUILD_CUDA} ]; then
@@ -114,10 +116,7 @@ fi
 if [ ! -z ${GG_BUILD_VULKAN} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_VULKAN=1"
 
-    # if on Mac, disable METAL
     if [[ "$OSTYPE" == "darwin"* ]]; then
-        CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=OFF -DGGML_BLAS=OFF"
-
         MACOS_RUNNER_CUSTOM_VULKAN_CMAKE_LOCATION="/usr/local/lib/cmake/vulkan"
         MACOS_RUNNER_CUSTOM_SPIRV_HEADERS_LOCATION="${MACOS_RUNNER_CUSTOM_VULKAN_CMAKE_LOCATION}/SPIRV-Headers/SPIRV-HeadersConfig.cmake"
         if [[ -f "${MACOS_RUNNER_CUSTOM_SPIRV_HEADERS_LOCATION}" || -h "${MACOS_RUNNER_CUSTOM_SPIRV_HEADERS_LOCATION}" ]]; then
@@ -133,7 +132,7 @@ if [ ! -z ${GG_BUILD_VULKAN} ]; then
 fi
 
 if [ ! -z ${GG_BUILD_WEBGPU} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_WEBGPU=1 -DGGML_METAL=OFF -DGGML_BLAS=OFF"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_WEBGPU=1"
 
     if [ ! -z "${GG_BUILD_WEBGPU_DAWN_PREFIX}" ]; then
         if [ -z "${CMAKE_PREFIX_PATH}" ]; then
@@ -167,6 +166,8 @@ fi
 
 if [ ! -z ${GG_BUILD_BLAS} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=${GG_BUILD_BLAS_VENDOR:-OpenBLAS}"
+else
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_BLAS=OFF"
 fi
 
 if [ ! -z ${GG_BUILD_OPENVINO} ]; then
@@ -700,8 +701,8 @@ function gg_sum_test_backend_ops_cpu {
 
 ## main
 
-export LLAMA_LOG_PREFIX=1
-export LLAMA_LOG_TIMESTAMPS=1
+export LLAMA_ARG_LOG_PREFIX=1
+export LLAMA_ARG_LOG_TIMESTAMPS=1
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
     # Create symlink: ./llama.cpp/models-mnt -> $MNT/models

common/arg.cpp

@@ -50,8 +50,6 @@
 
 #define LLAMA_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
 
-extern const char * LICENSES[];
-
 using json = nlohmann::ordered_json;
 using namespace common_arg_utils;
 
@@ -342,9 +340,7 @@ struct handle_model_result {
 };
 
 static handle_model_result common_params_handle_model(struct common_params_model & model,
-                                                      const std::string          & bearer_token,
-                                                      bool                         offline,
-                                                      bool                         search_mtp = false) {
+                                                      const common_download_opts & opts) {
     handle_model_result result;
 
     if (!model.docker_repo.empty()) {
@@ -356,10 +352,9 @@ static handle_model_result common_params_handle_model(struct common_params_model
             model.hf_file = model.path;
             model.path = "";
         }
-        common_download_opts opts;
-        opts.bearer_token = bearer_token;
-        opts.offline = offline;
-        auto download_result = common_download_model(model, opts, true, search_mtp);
+        common_download_opts hf_opts = opts;
+        hf_opts.download_mmproj = true; // also look for mmproj when downloading hf model
+        auto download_result = common_download_model(model, hf_opts);
 
         if (download_result.model_path.empty()) {
             throw std::runtime_error("failed to download model from Hugging Face");
@@ -384,9 +379,6 @@ static handle_model_result common_params_handle_model(struct common_params_model
             model.path = fs_get_cache_file(string_split<std::string>(f, '/').back());
         }
 
-        common_download_opts opts;
-        opts.bearer_token = bearer_token;
-        opts.offline = offline;
         auto download_result = common_download_model(model, opts);
         if (download_result.model_path.empty()) {
             throw std::runtime_error("failed to download model from " + model.url);
@@ -443,35 +435,49 @@ static bool parse_bool_value(const std::string & value) {
 // CLI argument parsing functions
 //
 
-void common_params_handle_models(common_params & params, llama_example curr_ex) {
+bool common_params_handle_models(common_params & params, llama_example curr_ex) {
     const bool spec_type_draft_mtp = std::find(params.speculative.types.begin(),
                                          params.speculative.types.end(),
                                          COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
 
-    auto res = common_params_handle_model(params.model, params.hf_token, params.offline, spec_type_draft_mtp);
-    if (params.no_mmproj) {
-        params.mmproj = {};
-    } else if (res.found_mmproj && params.mmproj.path.empty() && params.mmproj.url.empty()) {
-        // optionally, handle mmproj model when -hf is specified
-        params.mmproj = res.mmproj;
-    }
-    // only download mmproj if the current example is using it
-    for (const auto & ex : mmproj_examples) {
-        if (curr_ex == ex) {
-            common_params_handle_model(params.mmproj,    params.hf_token, params.offline);
-            break;
+    common_download_opts opts;
+    opts.bearer_token  = params.hf_token;
+    opts.offline       = params.offline;
+    opts.skip_download = params.skip_download;
+    opts.download_mtp  = spec_type_draft_mtp;
+
+    try {
+        auto res = common_params_handle_model(params.model, opts);
+        if (params.no_mmproj) {
+            params.mmproj = {};
+        } else if (res.found_mmproj && params.mmproj.path.empty() && params.mmproj.url.empty()) {
+            // optionally, handle mmproj model when -hf is specified
+            params.mmproj = res.mmproj;
         }
+        // only download mmproj if the current example is using it
+        for (const auto & ex : mmproj_examples) {
+            if (curr_ex == ex) {
+                common_params_handle_model(params.mmproj, opts);
+                break;
+            }
+        }
+
+        // when --spec-type mtp is set and no draft model was provided explicitly,
+        // fall back to the MTP head discovered alongside the -hf model
+        if (spec_type_draft_mtp && res.found_mtp &&
+            params.speculative.draft.mparams.path.empty() &&
+            params.speculative.draft.mparams.hf_repo.empty() &&
+            params.speculative.draft.mparams.url.empty()) {
+            params.speculative.draft.mparams.path = res.mtp.path;
+        }
+        common_params_handle_model(params.speculative.draft.mparams, opts);
+        common_params_handle_model(params.vocoder.model,             opts);
+        return true;
+    } catch (const common_skip_download_exception &) {
+        return false;
+    } catch (const std::exception &) {
+        throw;
     }
-    // when --spec-type mtp is set and no draft model was provided explicitly,
-    // fall back to the MTP head discovered alongside the -hf model
-    if (spec_type_draft_mtp && res.found_mtp &&
-        params.speculative.draft.mparams.path.empty() &&
-        params.speculative.draft.mparams.hf_repo.empty() &&
-        params.speculative.draft.mparams.url.empty()) {
-        params.speculative.draft.mparams.path = res.mtp.path;
-    }
-    common_params_handle_model(params.speculative.draft.mparams, params.hf_token, params.offline);
-    common_params_handle_model(params.vocoder.model,             params.hf_token, params.offline);
 }
 
 static bool common_params_parse_ex(int argc, char ** argv, common_params_context & ctx_arg) {
@@ -1091,16 +1097,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             exit(0);
         }
     ));
-    add_opt(common_arg(
-        {"--license"},
-        "show source code license and dependencies",
-        [](common_params &) {
-            for (int i = 0; LICENSES[i]; ++i) {
-                printf("%s\n", LICENSES[i]);
-            }
-            exit(0);
-        }
-    ));
     add_opt(common_arg(
         {"-cl", "--cache-list"},
         "show list of models in cache",
@@ -2998,7 +2994,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
             key_file.close();
         }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_API_KEY_FILE"));
     add_opt(common_arg(
         {"--ssl-key-file"}, "FNAME",
         "path to file a PEM-encoded SSL private key",
@@ -3026,7 +3022,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                 params.default_template_kwargs[item.key()] = item.value().dump();
             }
         }
-    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_CHAT_TEMPLATE_KWARGS"));
+    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_CHAT_TEMPLATE_KWARGS"));
     add_opt(common_arg(
         {"-to", "--timeout"}, "N",
         string_format("server read/write timeout in seconds (default: %d)", params.timeout_read),
@@ -3327,7 +3323,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params &, const std::string & value) {
             common_log_set_file(common_log_main(), value.c_str());
         }
-    ).set_env("LLAMA_LOG_FILE"));
+    ).set_env("LLAMA_ARG_LOG_FILE"));
     add_opt(common_arg(
         {"--log-colors"}, "[on|off|auto]",
         "Set colored logging ('on', 'off', or 'auto', default: 'auto')\n"
@@ -3344,7 +3340,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                     string_format("error: unknown value for --log-colors: '%s'\n", value.c_str()));
             }
         }
-    ).set_env("LLAMA_LOG_COLORS"));
+    ).set_env("LLAMA_ARG_LOG_COLORS"));
     add_opt(common_arg(
         {"-v", "--verbose", "--log-verbose"},
         "Set verbosity level to infinity (i.e. log all messages, useful for debugging)",
@@ -3359,7 +3355,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.offline = true;
         }
-    ).set_env("LLAMA_OFFLINE"));
+    ).set_env("LLAMA_ARG_OFFLINE"));
     add_opt(common_arg(
         {"-lv", "--verbosity", "--log-verbosity"}, "N",
         string_format("Set the verbosity threshold. Messages with a higher verbosity will be ignored. Values:\n"
@@ -3374,7 +3370,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.verbosity = value;
             common_log_set_verbosity_thold(value);
         }
-    ).set_env("LLAMA_LOG_VERBOSITY"));
+    ).set_env("LLAMA_ARG_LOG_VERBOSITY"));
     add_opt(common_arg(
         {"--log-prefix"},
         {"--no-log-prefix"},

common/arg.h

@@ -129,8 +129,11 @@ bool common_params_to_map(int argc, char ** argv, llama_example ex, std::map<com
 // see: https://github.com/ggml-org/llama.cpp/issues/18163
 void common_params_add_preset_options(std::vector<common_arg> & args);
 
-// Populate model paths (main model, mmproj, etc) from -hf if necessary
-void common_params_handle_models(common_params & params, llama_example curr_ex);
+// populate model paths (main model, mmproj, etc) from -hf if necessary
+// return true if the model is ready to use
+// throw an exception if there is an error that prevents the model from being used (e.g. network error, model not found, etc)
+// if params.skip_download is true, no downloads will be attempted. return false if the model is invalid or missing (e.g. ETag check failed)
+bool common_params_handle_models(common_params & params, llama_example curr_ex);
 
 // initialize argument parser context - used by test-arg-parser and preset
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);

common/common.h

@@ -479,7 +479,7 @@ struct common_params {
 
     std::set<std::string> model_alias;     // model aliases                                                 // NOLINT
     std::set<std::string> model_tags;      // model tags (informational, not used for routing)              // NOLINT
-    std::string hf_token             = ""; // HF token                                                      // NOLINT
+    std::string hf_token             = ""; // HF token (aka bearer token)                                   // NOLINT
     std::string prompt               = "";                                                                  // NOLINT
     std::string system_prompt        = "";                                                                  // NOLINT
     std::string prompt_file          = ""; // store the external prompt file name                           // NOLINT
@@ -507,6 +507,7 @@ struct common_params {
     int32_t control_vector_layer_start = -1; // layer range for control vector
     int32_t control_vector_layer_end   = -1; // layer range for control vector
     bool    offline                    = false;
+    bool    skip_download              = false; // skip model file downloading
 
     int32_t ppl_stride      = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
     int32_t ppl_output_type = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
@@ -587,7 +588,7 @@ struct common_params {
     // server params
     int32_t port                = 8080;          // server listens on this network port
     bool    reuse_port          = false;         // allow multiple sockets to bind to the same port
-    int32_t timeout_read        = 600;           // http read timeout in seconds
+    int32_t timeout_read        = 3600;          // http read timeout in seconds
     int32_t timeout_write       = timeout_read;  // http write timeout in seconds
     int32_t n_threads_http      = -1;    // number of threads to process HTTP requests (TODO: support threadpool)
     int32_t n_cache_reuse       = 0;     // min chunk size to reuse from the cache via KV shifting

common/download.cpp

@@ -292,6 +292,10 @@ static int common_download_file_single_online(const std::string & url,
 
     const bool file_exists = std::filesystem::exists(path);
 
+    if (!file_exists && opts.skip_download) {
+        return -2; // file is missing and download is disabled
+    }
+
     if (file_exists && skip_etag) {
         LOG_DBG("%s: using cached file: %s\n", __func__, path.c_str());
         return 304; // 304 Not Modified - fake cached response
@@ -357,6 +361,10 @@ static int common_download_file_single_online(const std::string & url,
             LOG_DBG("%s: using cached file (same etag): %s\n", __func__, path.c_str());
             return 304; // 304 Not Modified - fake cached response
         }
+        // pass this point, the file exists but is different from the server version, so we need to redownload it
+        if (opts.skip_download) {
+            return -2; // special code to indicate that the download was skipped due to etag mismatch
+        }
         if (remove(path.c_str()) != 0) {
             LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
             return -1;
@@ -775,13 +783,13 @@ static std::vector<download_task> get_url_tasks(const common_params_model & mode
 }
 
 common_download_model_result common_download_model(const common_params_model  & model,
-                                                   const common_download_opts & opts,
-                                                   bool download_mmproj,
-                                                   bool download_mtp) {
+                                                   const common_download_opts & opts) {
     common_download_model_result result;
     std::vector<download_task> tasks;
     hf_plan hf;
 
+    bool download_mmproj = opts.download_mmproj;
+    bool download_mtp = opts.download_mtp;
     bool is_hf = !model.hf_repo.empty();
 
     if (is_hf) {
@@ -806,18 +814,22 @@ common_download_model_result common_download_model(const common_params_model  &
         return result;
     }
 
-    std::vector<std::future<bool>> futures;
+    std::vector<std::future<int>> futures;
     for (const auto & task : tasks) {
         futures.push_back(std::async(std::launch::async,
             [&task, &opts, is_hf]() {
-                int status = common_download_file_single(task.url, task.path, opts, is_hf);
-                return is_http_status_ok(status);
+                return common_download_file_single(task.url, task.path, opts, is_hf);
             }
         ));
     }
 
     for (auto & f : futures) {
-        if (!f.get()) {
+        int status = f.get();
+        if (status == -2 && opts.skip_download) {
+            throw common_skip_download_exception();
+        }
+        bool is_ok = is_http_status_ok(status);
+        if (!is_ok) {
             return {};
         }
     }

common/download.h

@@ -52,6 +52,9 @@ struct common_download_opts {
     std::string bearer_token;
     common_header_list headers;
     bool offline = false;
+    bool skip_download = false; // if true, only validation is performed, common_skip_download_exception may be thrown if the file is missing or invalid
+    bool download_mmproj = false;
+    bool download_mtp = false;
     common_download_callback * callback = nullptr;
 };
 
@@ -62,6 +65,11 @@ struct common_download_model_result {
     std::string mtp_path;
 };
 
+// throw if the file is missing or invalid (e.g. ETag check failed)
+struct common_skip_download_exception : public std::runtime_error {
+    common_skip_download_exception() : std::runtime_error("skip download") {}
+};
+
 // Download model from HuggingFace repo or URL
 //
 // input (via model struct):
@@ -89,9 +97,7 @@ struct common_download_model_result {
 // returns result with model_path, mmproj_path and mtp_path (empty when not found / on failure)
 common_download_model_result common_download_model(
     const common_params_model & model,
-    const common_download_opts & opts = {},
-    bool download_mmproj = false,
-    bool download_mtp    = false
+    const common_download_opts & opts = {}
 );
 
 // returns list of cached models
@@ -99,6 +105,7 @@ std::vector<common_cached_model_info> common_list_cached_models();
 
 // download single file from url to local path
 // returns status code or -1 on error
+// returns -2 if the download was skipped due to ETag mismatch (file outdated, skip_download=true)
 // skip_etag: if true, don't read/write .etag files (for HF cache where filename is the hash)
 int common_download_file_single(const std::string & url,
                                 const std::string & path,

common/ngram-mod.cpp

@@ -1,5 +1,7 @@
 #include "ngram-mod.h"
 
+#include <algorithm>
+
 //
 // common_ngram_mod
 //

conversion/__init__.py

@@ -47,6 +47,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "DeepseekForCausalLM": "deepseek",
     "DeepseekV2ForCausalLM": "deepseek",
     "DeepseekV3ForCausalLM": "deepseek",
+    "DeepseekV32ForCausalLM": "deepseek",
     "DistilBertForMaskedLM": "bert",
     "DistilBertForSequenceClassification": "bert",
     "DistilBertModel": "bert",
@@ -74,6 +75,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "Gemma3nForCausalLM": "gemma",
     "Gemma3nForConditionalGeneration": "gemma",
     "Gemma4ForConditionalGeneration": "gemma",
+    "Gemma4ForCausalLM": "gemma",
     "GemmaForCausalLM": "gemma",
     "Glm4ForCausalLM": "glm",
     "Glm4MoeForCausalLM": "glm",
@@ -215,6 +217,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "T5EncoderModel": "t5",
     "T5ForConditionalGeneration": "t5",
     "T5WithLMHeadModel": "t5",
+    "TalkieForCausalLM": "talkie",
     "UMT5ForConditionalGeneration": "t5",
     "UMT5Model": "t5",
     "UltravoxModel": "ultravox",
@@ -234,6 +237,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
 MMPROJ_MODEL_MAP: dict[str, str] = {
     "AudioFlamingo3ForConditionalGeneration": "ultravox",
     "CogVLMForCausalLM": "cogvlm",
+    "DeepseekOCR2ForCausalLM": "deepseek",
     "DeepseekOCRForCausalLM": "deepseek",
     "DotsOCRForCausalLM": "dotsocr",
     "Gemma3ForConditionalGeneration": "gemma",

conversion/base.py

@@ -119,7 +119,8 @@ class ModelBase:
                  small_first_shard: bool = False, hparams: dict[str, Any] | None = None, remote_hf_model_id: str | None = None,
                  disable_mistral_community_chat_template: bool = False,
                  sentence_transformers_dense_modules: bool = False,
-                 fuse_gate_up_exps: bool = False):
+                 fuse_gate_up_exps: bool = False,
+                 fp8_as_q8: bool = False):
         if type(self) is ModelBase or \
                 type(self) is TextModel or \
                 type(self) is MmprojModel:
@@ -148,6 +149,8 @@ class ModelBase:
         self.dir_model_card = dir_model  # overridden in convert_lora_to_gguf.py
         self._is_nvfp4 = False
         self._is_mxfp4 = False
+        self._fp8_as_q8 = fp8_as_q8
+        self._fp8_dequantized: set[str] = set()
 
         # Apply heuristics to figure out typical tensor encoding based on first tensor's dtype
         # NOTE: can't use field "torch_dtype" in config.json, because some finetunes lie.
@@ -429,6 +432,8 @@ class ModelBase:
                         s = self.model_tensors[name]
                         self.model_tensors[weight_name] = lambda w=w, s=s, bs=block_size: dequant_simple(w(), s(), bs)
                         tensors_to_remove.append(name)
+                        if self._fp8_as_q8:
+                            self._fp8_dequantized.add(weight_name)
                     if name.endswith(".activation_scale"):  # unused
                         tensors_to_remove.append(name)
                     if name.endswith("_activation_scale"):  # Mistral-Small-4-119B-2602, unused
@@ -440,6 +445,8 @@ class ModelBase:
                         s = self.model_tensors[name]
                         self.model_tensors[weight_name] = lambda w=w, s=s, bs=block_size: dequant_simple(w(), s(), bs)
                         tensors_to_remove.append(name)
+                        if self._fp8_as_q8:
+                            self._fp8_dequantized.add(weight_name)
                     if name.endswith(".qscale_act"):
                         tensors_to_remove.append(name)
             elif quant_method == "gptq":
@@ -467,7 +474,14 @@ class ModelBase:
             elif quant_method == "compressed-tensors":
                 quant_format = quant_config["format"]
                 groups = quant_config["config_groups"]
-                if len(groups) > 1:
+                nvfp4_compressed_tensors = (
+                    quant_format == "nvfp4-pack-quantized"
+                    or quant_format == "mixed-precision"
+                    and bool(groups)
+                    and all(g.get("format") == "nvfp4-pack-quantized" for g in groups.values() if isinstance(g, dict))
+                )
+
+                if len(groups) > 1 and not nvfp4_compressed_tensors:
                     raise NotImplementedError("Can't handle multiple config groups for compressed-tensors yet")
                 weight_config = tuple(groups.values())[0]["weights"]
 
@@ -476,6 +490,11 @@ class ModelBase:
                     strategy = weight_config.get("strategy")
                     assert strategy == "channel" or strategy == "block"
                     assert weight_config.get("group_size") is None  # didn't find a model using this yet
+                    is_fp8 = (
+                        quant_format == "float-quantized"
+                        and weight_config.get("type") == "float"
+                        and weight_config.get("num_bits") == 8
+                    )
                     for name in self.model_tensors.keys():
                         if name.endswith(".weight_scale"):
                             weight_name = name.removesuffix("_scale")
@@ -483,6 +502,8 @@ class ModelBase:
                             s = self.model_tensors[name]
                             self.model_tensors[weight_name] = lambda w=w, s=s: dequant_simple(w(), s(), block_size)
                             tensors_to_remove.append(name)
+                            if self._fp8_as_q8 and is_fp8:
+                                self._fp8_dequantized.add(weight_name)
                 elif quant_format == "pack-quantized":
                     assert weight_config.get("strategy") == "group"
                     assert weight_config.get("type", "int") == "int"
@@ -505,6 +526,9 @@ class ModelBase:
                             tensors_to_remove += [base_name + n for n in ("_packed", "_shape", "_scale")]
                             if (base_name + "_zero_point") in self.model_tensors:
                                 tensors_to_remove.append(base_name + "_zero_point")
+                elif nvfp4_compressed_tensors:
+                    # Don't error from compressed-tensors, we'll handle them in _generate_nvfp4_tensors
+                    pass
                 else:
                     raise NotImplementedError(f"Quant format {quant_format!r} for method {quant_method!r} is not yet supported")
             elif quant_method == "modelopt":
@@ -514,10 +538,18 @@ class ModelBase:
                 for name in self.model_tensors.keys():
                     if name.endswith(".weight_scale"):
                         weight_name = name.removesuffix("_scale")
+                        if weight_name not in self.model_tensors:
+                            tensors_to_remove.append(name)
+                            continue
                         w = self.model_tensors[weight_name]
                         s = self.model_tensors[name]
+                        is_fp8_weight = False
+                        if self._fp8_as_q8:
+                            is_fp8_weight = w().dtype in (torch.float8_e4m3fn, torch.float8_e5m2)
                         self.model_tensors[weight_name] = lambda w=w, s=s: dequant_simple(w(), s(), None)
                         tensors_to_remove.append(name)
+                        if is_fp8_weight:
+                            self._fp8_dequantized.add(weight_name)
                     if name.endswith((".input_scale", ".k_scale", ".v_scale")):
                         tensors_to_remove.append(name)
             elif quant_method is not None:
@@ -605,8 +637,10 @@ class ModelBase:
         return [(new_name, data_torch)]
 
     def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
-        del name, new_name, bid, n_dims  # unused
-
+        del new_name, bid  # unused
+        # Force FP8-original tensors to Q8_0 when requested; Q8_0 is faster than F16/BF16.
+        if self._fp8_as_q8 and name in self._fp8_dequantized and n_dims >= 2:
+            return gguf.GGMLQuantizationType.Q8_0
         return False
 
     # some models need extra generated tensors (like rope_freqs)
@@ -746,10 +780,13 @@ class ModelBase:
         del experts, merged
 
     def prepare_tensors(self):
-        # detect NVFP4 quantization (ModelOpt format)
-        quant_algo = (self.hparams.get("quantization_config") or {}).get("quant_algo")
-        quant_method = (self.hparams.get("quantization_config") or {}).get("quant_method")
-        quant_layers = (self.hparams.get("quantization_config") or {}).get("quantized_layers") or {}
+        # detect NVFP4 quantization (ModelOpt and Compressed-tensors formats)
+        quantization_config = self.hparams.get("quantization_config") or {}
+        quant_algo = quantization_config.get("quant_algo")
+        quant_method = quantization_config.get("quant_method")
+        quant_format = quantization_config.get("format")
+        quant_groups = quantization_config.get("config_groups") or {}
+        quant_layers = quantization_config.get("quantized_layers") or {}
         quant_config_file = self.dir_model / "hf_quant_config.json"
 
         if (not quant_algo or not quant_layers) and quant_config_file.is_file():
@@ -760,13 +797,25 @@ class ModelBase:
                 producer_name = (producer.get("name") or "").lower()
                 if quant_method is None:
                     self.hparams.setdefault("quantization_config", {})["quant_method"] = producer_name
+                    quant_method = producer_name
                 quant_algo = quant_config.get("quant_algo", quant_algo)
+                quant_method = quant_config.get("quant_method", quant_method)
+                quant_format = quant_config.get("format", quant_format)
+                quant_groups = quant_config.get("config_groups", quant_groups) or {}
                 quant_layers = quant_config.get("quantized_layers", quant_layers) or {}
 
         # Some models use per-tensor quant_algo (e.g. "MIXED_PRECISION" with
         # per-layer NVFP4/FP8) instead of a single global "NVFP4" value.
+        nvfp4_compressed_tensors = quant_method == "compressed-tensors" and (
+            quant_format == "nvfp4-pack-quantized"
+            or quant_format == "mixed-precision"
+            and bool(quant_groups)
+            and all(g.get("format") == "nvfp4-pack-quantized" for g in quant_groups.values() if isinstance(g, dict))
+        )
         if quant_algo != "NVFP4":
-            if any(v.get("quant_algo") == "NVFP4" for v in quant_layers.values() if isinstance(v, dict)):
+            if nvfp4_compressed_tensors:
+                quant_algo = "NVFP4"
+            elif any(str(v.get("quant_algo")).endswith("NVFP4") for v in quant_layers.values() if isinstance(v, dict)):
                 quant_algo = "NVFP4"
 
         self._is_nvfp4 = quant_algo == "NVFP4"
@@ -776,6 +825,28 @@ class ModelBase:
         # This must run before dequant_model so NVFP4 tensors are removed
         # from model_tensors, leaving only non-NVFP4 (e.g. FP8) for dequant.
         if self._is_nvfp4:
+            if nvfp4_compressed_tensors:
+                # Convert compressed-tensors 'global' scales into the reciprocal
+                def inverse_scale(gen):
+                    def load():
+                        scale = LazyTorchTensor.to_eager(gen()).float()
+                        return 1.0 / scale
+                    return load
+
+                # Change the compressed-tensors names to the ModelOpt names for handling consistently later
+                for name in list(self.model_tensors.keys()):
+                    if name.endswith(".weight_packed"):
+                        weight_name = name.removesuffix("_packed")
+                        if weight_name not in self.model_tensors:
+                            self.model_tensors[weight_name] = self.model_tensors.pop(name)
+                    elif name.endswith(".weight_global_scale"):
+                        scale2_name = name.replace(".weight_global_scale", ".weight_scale_2")
+                        if scale2_name not in self.model_tensors:
+                            self.model_tensors[scale2_name] = inverse_scale(self.model_tensors.pop(name))
+                    elif name.endswith(".input_global_scale"):
+                        input_scale_name = name.replace(".input_global_scale", ".input_scale")
+                        if input_scale_name not in self.model_tensors:
+                            self.model_tensors[input_scale_name] = inverse_scale(self.model_tensors.pop(name))
             self._generate_nvfp4_tensors()
 
         self.dequant_model()
@@ -844,6 +915,8 @@ class ModelBase:
                             gguf.MODEL_TENSOR.SSM_CONV1D_Q,
                             gguf.MODEL_TENSOR.SSM_CONV1D_K,
                             gguf.MODEL_TENSOR.SSM_CONV1D_V,
+                            # DSA indexer weights should be F32
+                            gguf.MODEL_TENSOR.INDEXER_PROJ,
                         )
                     )
                     or new_name[-7:] not in (".weight", ".lora_a", ".lora_b")
@@ -1067,7 +1140,7 @@ class TextModel(ModelBase):
         # Skip multimodal tensors
         if name.startswith(("mlp", "vit.", "vpm.", "siglip2.", "conformer.", "merger.", "resampler.", "sound_encoder.", "sound_projection.", "speech_embeddings.")) \
                 or "visual." in name or "vision." in name or "audio." in name or "talker." in name \
-                or "vision_" in name or "audio_" in name or "sam_model" in name \
+                or "vision_" in name or "audio_" in name \
                 or "token2wav." in name or "code2wav." in name \
                 or "projector." in name or "pre_mm_projector_norm" in name \
                 or "image_newline" in name or "view_seperator" in name \
@@ -1374,6 +1447,9 @@ class TextModel(ModelBase):
         if chkhsh == "0fe1cf6eda062318a1af7270f3331a85c539a01778ff948e24388e949c5282f4":
             # ref: https://huggingface.co/evilfreelancer/ruGPT3XL
             res = "gpt-2"
+        if chkhsh == "9e454714343b69b99b71795c1d27a68c2a1d15dab111f4d353109f966af29da7":
+            # ref: https://huggingface.co/LiquidAI/LFM2.5-8B-A1B
+            res = "lfm2"
         if chkhsh == "0ef9807a4087ebef797fc749390439009c3b9eda9ad1a097abbe738f486c01e5":
             # ref: https://huggingface.co/meta-llama/Meta-Llama-3-8B
             res = "llama-bpe"
@@ -1525,7 +1601,7 @@ class TextModel(ModelBase):
             # ref: https://huggingface.co/K-intelligence/Midm-2.0-Base-Instruct
             res = "midm-2.0"
         if chkhsh == "169bf0296a13c4d9b7672313f749eb36501d931022de052aad6e36f2bf34dd51":
-            # ref: https://huggingface.co/LiquidAI/LFM2-Tokenizer
+            # ref: https://huggingface.co/LiquidAI/LFM2.5-350M
             res = "lfm2"
         if chkhsh == "2085e1638f6c377a0aa4ead21b27bb4cb941bf800df86ed391011769c1758dfb":
             # ref: https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B
@@ -1575,6 +1651,12 @@ class TextModel(ModelBase):
         if chkhsh == "62f6fb0a6fd5098caeabb19b07a5c1099cafc8b9c40eab6ea89ece4ec02fbc57":
             # ref: https://huggingface.co/sarvamai/sarvam-30b
             res = "sarvam-moe"
+        if chkhsh == "f728162c1315c26e40249849799b4ba3fe584c32084b4795b03eb295e63cb5af":
+            # ref: https://huggingface.co/lewtun/talkie-1930-13b-it-hf
+            res = "talkie"
+        if chkhsh == "36f3066e97b7f3994b379aaacde306c1444c6ae84e81a5ae3cd2b7ed3b8c42d4":
+            # ref: https://huggingface.co/openbmb/MiniCPM5-1B
+            res = "minicpm5"
 
         if res is None:
             logger.warning("\n")
@@ -1610,6 +1692,16 @@ class TextModel(ModelBase):
         special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
         special_vocab.add_to_gguf(self.gguf_writer)
 
+    def _set_vocab_whitespace(self) -> None:
+        tokens, toktypes, _ = self.get_vocab_base()
+        self.gguf_writer.add_tokenizer_model("whitespace")
+        self.gguf_writer.add_tokenizer_pre("whitespace") # pinned, not hash-detected: chktxt hash collides with jina-v1-en
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        special_vocab.add_to_gguf(self.gguf_writer)
+
     def _set_vocab_hybriddna(self):
         from transformers import AutoTokenizer
         tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
@@ -2364,10 +2456,9 @@ class MmprojModel(ModelBase):
         raise KeyError(f"could not find any of: {keys}")
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
-        del bid, name, n_dims  # unused
         if ".patch_embd.weight" in new_name or ".patch_merger.weight" in new_name:
             return gguf.GGMLQuantizationType.F16 if self.ftype == gguf.LlamaFileType.MOSTLY_F16 else gguf.GGMLQuantizationType.F32
-        return False
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
 
 
 class LazyTorchTensor(gguf.LazyBase):

conversion/bert.py

@@ -571,7 +571,16 @@ class JinaBertV2Model(BertModel):
         if tokenizer_class == 'BertTokenizer':
             super().set_vocab()
         elif tokenizer_class == 'RobertaTokenizer':
-            self._set_vocab_gpt2()
+            pre_tokenizer_type = None
+            tokenizer_json_path = self.dir_model / "tokenizer.json"
+            if tokenizer_json_path.is_file():
+                with open(tokenizer_json_path, "r", encoding="utf-8") as f:
+                    pre_tokenizer_type = json.load(f).get("pre_tokenizer", {}).get("type")
+
+            if pre_tokenizer_type == "Whitespace":
+                self._set_vocab_whitespace()
+            else:
+                self._set_vocab_gpt2()
             self.gguf_writer.add_token_type_count(2)
         else:
             raise NotImplementedError(f'Tokenizer {tokenizer_class} is not supported for JinaBertModel')

conversion/deepseek.py

@@ -16,10 +16,14 @@ from .qwen import QwenModel
 
 @ModelBase.register("DeepseekOCRForCausalLM")
 class DeepseekOCRVisionModel(MmprojModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.clip_projector_type = gguf.VisionProjectorType.DEEPSEEKOCR
+
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
-        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.DEEPSEEKOCR)
+        self.gguf_writer.add_clip_projector_type(self.clip_projector_type)
         # default values below are taken from HF tranformers code
         self.gguf_writer.add_vision_attention_layernorm_eps(hparams.get("layer_norm_eps", 1e-6))
         self.gguf_writer.add_vision_use_gelu(True)
@@ -49,22 +53,27 @@ class DeepseekOCRVisionModel(MmprojModel):
             raise ValueError("DeepseekOCR model requires 'vision_config' in the model configuration, but it was not found")
 
         vision_config['sam'] = vision_config['width']['sam_vit_b']
-        vision_config.update(vision_config['width']['clip-l-14-224'])
-        vision_config['hidden_size'] = vision_config['width']
-        vision_config['num_heads'] = vision_config['heads']
-        vision_config['intermediate_size'] = vision_config['heads'] * 4
+        if vision_config['width'].get('clip-l-14-224') is not None:
+            vision_config.update(vision_config['width']['clip-l-14-224'])
+        if isinstance(vision_config['width'], int):
+            vision_config['hidden_size'] = vision_config['width']
+        if vision_config.get('heads') is not None:
+            vision_config['num_heads'] = vision_config['heads']
+            vision_config['intermediate_size'] = vision_config['heads'] * 4
 
         return vision_config
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
-        if ".embeddings." in name or 'pos_embed' in name:
-            return gguf.GGMLQuantizationType.F32
-        if ".rel_pos_h" in name or '.rel_pos_w' in name:
-            return gguf.GGMLQuantizationType.F32
-        if ".neck." in name or ".net_" in name:
-            return gguf.GGMLQuantizationType.F32
+        for nq_name in ('.embeddings.', 'pos_embed', '.rel_pos_h', '.rel_pos_w', '.neck.', '.net_'):
+            if nq_name in name:
+                return gguf.GGMLQuantizationType.F32
         return super().tensor_force_quant(name, new_name, bid, n_dims)
 
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.endswith("view_seperator"):
+            data_torch = data_torch.unsqueeze(0)
+        yield from super().modify_tensors(data_torch, name, bid)
+
     @classmethod
     def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
         name, gen = item
@@ -81,6 +90,33 @@ class DeepseekOCRVisionModel(MmprojModel):
         return super().filter_tensors((name, gen))
 
 
+@ModelBase.register("DeepseekOCR2ForCausalLM")
+class DeepseekOCR2VisionModel(DeepseekOCRVisionModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.clip_projector_type = gguf.VisionProjectorType.DEEPSEEKOCR2
+
+    def set_gguf_parameters(self):
+        # the vision tower's qwen2 encoder is built from fixed defaults,
+        # see build_qwen2_decoder_as_encoder() in deepencoderv2.py
+        if self.hparams.get("patch_size") is None:
+            self.hparams["patch_size"] = 16
+        if self.hparams.get("intermediate_size") is None:
+            self.hparams["intermediate_size"] = 4864
+        if self.hparams.get("num_attention_heads") is None:
+            self.hparams["num_attention_heads"] = 14
+        super().set_gguf_parameters()
+        # qwen2 encoder is GQA: 14 Q heads, 2 KV heads
+        self.gguf_writer.add_vision_head_count_kv(2)
+
+    def get_vision_config(self) -> dict[str, Any]:
+        vision_config = super().get_vision_config()
+        vision_config['hidden_size'] = vision_config['width']['qwen2-0-5b']['dim']
+        if vision_config.get('layers') is None:
+            vision_config['layers'] = 24
+        return vision_config
+
+
 @ModelBase.register("DeepseekForCausalLM")
 class DeepseekModel(TextModel):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK
@@ -188,13 +224,21 @@ class DeepseekV2Model(TextModel):
         self.origin_hf_arch = hparams.get('architectures', [None])[0]
 
         # special handling for Deepseek OCR
-        if self.origin_hf_arch == "DeepseekOCRForCausalLM":
+        if self.origin_hf_arch in ("DeepseekOCRForCausalLM", "DeepseekOCR2ForCausalLM"):
             self.model_arch = gguf.MODEL_ARCH.DEEPSEEK2OCR
             self.gguf_writer.arch = gguf.MODEL_ARCH_NAMES[self.model_arch]
             self.gguf_writer.add_architecture()
             # default jinja template
             self.gguf_writer.add_chat_template("{% for m in messages %}{{m['content']}}{% endfor %}")
 
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, _ = item
+        # DeepSeek-OCR vision encoder (SAM + DeepSeek-OCR-2 qwen2 tower)
+        if "sam_model" in name or "qwen2_model" in name:
+            return None
+        return super().filter_tensors(item)
+
     def set_vocab(self):
         try:
             self._set_vocab_gpt2()
@@ -386,3 +430,32 @@ class DeepseekV2Model(TextModel):
             experts = [k for d in self._experts for k in d.keys()]
             if len(experts) > 0:
                 raise ValueError(f"Unprocessed experts: {experts}")
+
+
+@ModelBase.register("DeepseekV32ForCausalLM")
+class DeepseekV32Model(DeepseekV2Model):
+    model_arch = gguf.MODEL_ARCH.DEEPSEEK32
+    skip_mtp = False
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0)
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def set_vocab(self):
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
+        assert getattr(tokenizer, "add_bos_token", False), "Change value of add_bos_token to true in tokenizer_config.json file."
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        # NextN/MTP prediction layers
+        if (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
+            self.gguf_writer.add_nextn_predict_layers(num_nextn_predict_layers)
+
+        # DSA indexer parameters
+        self.gguf_writer.add_indexer_head_count(self.hparams["index_n_heads"])
+        self.gguf_writer.add_indexer_key_length(self.hparams["index_head_dim"])
+        self.gguf_writer.add_indexer_top_k(self.hparams["index_topk"])

conversion/gemma.py

@@ -614,7 +614,7 @@ class Gemma3NModel(Gemma3Model):
         yield from super().modify_tensors(data_torch, name, bid)
 
 
-@ModelBase.register("Gemma4ForConditionalGeneration")
+@ModelBase.register("Gemma4ForConditionalGeneration", "Gemma4ForCausalLM")
 class Gemma4Model(Gemma3Model):
     model_arch = gguf.MODEL_ARCH.GEMMA4
 
@@ -786,14 +786,15 @@ class Gemma4VisionAudioModel(MmprojModel):
         super().set_gguf_parameters()
 
         # vision params
+        assert self.hparams_vision is not None
         self.gguf_writer.add_clip_vision_projector_type(gguf.VisionProjectorType.GEMMA4V)
-        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams.get("layer_norm_eps", 1e-6))
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams_vision.get("layer_norm_eps", 1e-6))
 
         # audio params
-        if self.hparams_audio:
-            self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.GEMMA4A)
-            self.gguf_writer.add_audio_num_mel_bins(self.hparams_audio["feat_in"])
-            self.gguf_writer.add_audio_attention_layernorm_eps(1e-5)
+        assert self.hparams_audio is not None
+        self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.GEMMA4A)
+        self.gguf_writer.add_audio_num_mel_bins(self.hparams_audio["feat_in"])
+        self.gguf_writer.add_audio_attention_layernorm_eps(self.hparams_audio.get("layer_norm_eps", 1e-6))
 
     def is_audio_tensor(self, name: str) -> bool:
         return "audio_tower" in name or "embed_audio" in name

conversion/qwen.py

@@ -1,6 +1,5 @@
 from __future__ import annotations
 
-from pathlib import Path
 from typing import Any, Callable, Iterable, TYPE_CHECKING
 
 import torch
@@ -549,6 +548,7 @@ class _Qwen35MtpMixin:
     tensor_map: gguf.TensorNameMap
     no_mtp: bool
     mtp_only: bool
+    _original_block_count: int | None = None
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
@@ -557,22 +557,44 @@ class _Qwen35MtpMixin:
             self.block_count += self.hparams.get("mtp_num_hidden_layers", 0)
         self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
 
+    def index_tensors(self, remote_hf_model_id: str | None = None) -> dict[str, Callable[[], Tensor]]:
+        hparams = {**self.hparams, **self.hparams.get("text_config", {})}
+        key = next((k for k in ["n_layers", "num_hidden_layers", "n_layer", "num_layers"] if k in hparams), None)
+        type(self)._original_block_count = hparams.get(key)
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)  # ty: ignore[unresolved-attribute]
+
     @classmethod
     def filter_tensors(cls, item):
-        name, _ = item
+        assert cls._original_block_count is not None
+        # TODO: change TextModel to super()
+        if (titem := TextModel.filter_tensors(item)) is None:
+            return None
+        name, gen = titem
+        if name.startswith("model.mtp."):
+            name = name.replace("model.", "", 1)
         if name.startswith("mtp."):
             if cls.no_mtp:
                 return None
-            return item
-        if cls.mtp_only:
-            canonical = name.replace("language_model.", "")
-            keep = canonical in (
+            remapper = {
+                "fc":                    "eh_proj",
+                "pre_fc_norm_embedding": "enorm",
+                "pre_fc_norm_hidden":    "hnorm",
+                "norm":                  "shared_head.norm",
+            }
+            parts = name.split(".", 3)
+            if len(parts) == 4 and parts[1] == "layers" and parts[2].isdecimal():
+                mtp_idx = int(parts[2])
+                name = f"model.layers.{cls._original_block_count + mtp_idx}.{parts[3]}"
+            elif len(parts) == 3 and parts[1] in remapper:
+                name = f"model.layers.{cls._original_block_count}.{remapper[parts[1]]}.{parts[2]}"
+        elif cls.mtp_only:
+            keep = name in (
                 "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
                 "embed_tokens.weight", "norm.weight",
             )
             if not keep:
                 return None
-        return super().filter_tensors(item)  # ty: ignore[unresolved-attribute]
+        return name, gen
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()  # ty: ignore[unresolved-attribute]
@@ -594,29 +616,6 @@ class _Qwen35MtpMixin:
             self.metadata.version, size_label=None, output_type=output_type, model_type=None)    # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
         self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
 
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        if name.startswith("mtp."):
-            n_layer = self.hparams["num_hidden_layers"]
-            if name.find("layers.") != -1:
-                assert bid is not None
-                name = name.replace(f"mtp.layers.{bid}", f"model.layers.{bid + n_layer}")
-                bid = bid + n_layer
-            else:
-                remapper = {
-                    "mtp.fc":                    "model.layers.{bid}.eh_proj",
-                    "mtp.pre_fc_norm_embedding": "model.layers.{bid}.enorm",
-                    "mtp.pre_fc_norm_hidden":    "model.layers.{bid}.hnorm",
-                    "mtp.norm":                  "model.layers.{bid}.shared_head.norm",
-                }
-                stem   = Path(name).stem
-                suffix = Path(name).suffix
-                tmpl   = remapper[stem] + suffix
-                for b in range(n_layer, self.block_count):
-                    yield from super().modify_tensors(data_torch, tmpl.format(bid=b), b)  # ty: ignore[unresolved-attribute]
-                return
-
-        yield from super().modify_tensors(data_torch, name, bid)  # ty: ignore[unresolved-attribute]
-
 
 @ModelBase.register("Qwen3_5ForConditionalGeneration", "Qwen3_5ForCausalLM")
 class Qwen3_5TextModel(_Qwen35MtpMixin, _Qwen35MRopeMixin, _LinearAttentionVReorderBase):

conversion/talkie.py

@@ -0,0 +1,53 @@
+from __future__ import annotations
+
+from typing import Iterable, TYPE_CHECKING
+
+import torch
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import LazyTorchTensor, ModelBase, TextModel, gguf
+
+
+@ModelBase.register("TalkieForCausalLM")
+class TalkieModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.TALKIE
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        # Talkie used F.rms_norm without an explicit eps
+        self.gguf_writer.add_layer_norm_rms_eps(torch.finfo(torch.float32).eps)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        prefix = f"model.blocks.{bid}." if bid is not None else ""
+        suffix = name.removeprefix(prefix)
+
+        if suffix == "attn_gain.a_g":
+            yield self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_OUT, bid, ".scale"), data_torch
+            return
+        elif suffix == "mlp_gain.a_g":
+            yield self.format_tensor_name(gguf.MODEL_TENSOR.FFN_DOWN, bid, ".scale"), data_torch
+            return
+        elif suffix == "lm_head_gain.w_g":
+            self.gguf_writer.add_logit_scale(LazyTorchTensor.to_eager(data_torch).item())
+            return
+        elif suffix in ("attn.attn_query.weight", "attn.attn_key.weight"):
+            # absorb inverse rope
+            head_dim = self.hparams["head_dim"]
+            shape = data_torch.shape
+            data_torch = torch.reshape(data_torch, (-1, head_dim, shape[-1]))
+            signs = torch.ones((1, head_dim, 1), dtype=data_torch.dtype)
+            signs[:, head_dim // 2 :, :] = -1
+            if self.lazy:
+                signs = LazyTorchTensor.from_eager(signs)
+            # (n_head, head_dim, n_in) -> (n_out, n_in)
+            data_torch = torch.reshape(data_torch * signs, shape)
+        elif suffix == "attn.head_gain.head_g":
+            # allow head gain to broadcast
+            data_torch = data_torch.unsqueeze(-1)
+
+        if not name.endswith(".weight"):
+            name += ".weight"
+
+        yield from super().modify_tensors(data_torch, name, bid)

convert_hf_to_gguf.py

@@ -148,6 +148,10 @@ def parse_args() -> argparse.Namespace:
         "--fuse-gate-up-exps", action="store_true",
         help="Fuse gate_exps and up_exps tensors into a single gate_up_exps tensor for MoE models.",
     )
+    parser.add_argument(
+        "--fp8-as-q8", action="store_true",
+        help="Store tensors dequantized from FP8 as Q8_0 instead of BF16/F16.",
+    )
 
     args = parser.parse_args()
     if not args.print_supported_models and args.model is None:
@@ -264,7 +268,8 @@ def main() -> None:
                                      small_first_shard=args.no_tensor_first_split,
                                      remote_hf_model_id=hf_repo_id, disable_mistral_community_chat_template=disable_mistral_community_chat_template,
                                      sentence_transformers_dense_modules=args.sentence_transformers_dense_modules,
-                                     fuse_gate_up_exps=args.fuse_gate_up_exps
+                                     fuse_gate_up_exps=args.fuse_gate_up_exps,
+                                     fp8_as_q8=args.fp8_as_q8,
                                      )
 
         if args.vocab_only:

convert_hf_to_gguf_update.py

@@ -139,7 +139,7 @@ models = [
     {"name": "seed-coder",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base", },
     {"name": "a.x-4.0",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/skt/A.X-4.0", },
     {"name": "midm-2.0",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/K-intelligence/Midm-2.0-Base-Instruct", },
-    {"name": "lfm2",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2-Tokenizer"},
+    {"name": "lfm2",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2.5-350M", },
     {"name": "exaone4",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B", },
     {"name": "mellum",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/JetBrains/Mellum-4b-base", },
     {"name": "modern-bert",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/answerdotai/ModernBERT-base", },
@@ -156,6 +156,8 @@ models = [
     {"name": "kanana2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/kakaocorp/kanana-2-30b-a3b-instruct-2601", },
     {"name": "f2llmv2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/codefuse-ai/F2LLM-v2-4B", },
     {"name": "sarvam-moe",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sarvamai/sarvam-30b", },
+    {"name": "talkie",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/lewtun/talkie-1930-13b-it-hf", },
+    {"name": "minicpm5",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openbmb/MiniCPM5-1B"},
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions
@@ -181,6 +183,8 @@ pre_computed_hashes = [
     # jina-v2-de variants
     {"name": "jina-v2-de", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/aari1995/German_Semantic_V3", "chkhsh": "b3d1dd861f1d4c5c0d2569ce36baf3f90fe8a102db3de50dd71ff860d91be3df"},
     {"name": "gpt-2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/evilfreelancer/ruGPT3XL", "chkhsh": "0fe1cf6eda062318a1af7270f3331a85c539a01778ff948e24388e949c5282f4"},
+    # lfm2 variants
+    {"name": "lfm2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2.5-8B-A1B", "chkhsh": "9e454714343b69b99b71795c1d27a68c2a1d15dab111f4d353109f966af29da7"},
 ]
 
 

convert_lora_to_gguf.py

@@ -208,6 +208,16 @@ class LoraTorchTensor:
     def to(self, *args, **kwargs):
         return LoraTorchTensor(self._lora_A.to(*args, **kwargs), self._lora_B.to(*args, **kwargs))
 
+    def __mul__(self, other) -> LoraTorchTensor:
+        # Only output-side multiplication for now
+        # W = B @ A, so M_out * W == (M_out * B) @ A
+        if not isinstance(other, (int, float)) and other.shape and other.shape[-1] != 1:
+            raise NotImplementedError
+        return LoraTorchTensor(self._lora_A, self._lora_B * other)
+
+    def __rmul__(self, other) -> LoraTorchTensor:
+        return self * other
+
     @classmethod
     def __torch_function__(cls, func: Callable, types, args=(), kwargs=None):
         del types  # unused

docs/autoparser.md

@@ -459,7 +459,7 @@ Each returned parser is wrapped by `wrap_for_generation_prompt()`, which prepend
 
 - Usage: `./bin/llama-template-analysis path/to/template.jinja`
 
-**Debug Logging**: Enable with `LLAMA_LOG_VERBOSITY=2`
+**Debug Logging**: Enable with `LLAMA_ARG_LOG_VERBOSITY=2`
 
 - Shows detailed analysis steps, pattern extraction results, and generated parser structure
 

docs/backend/SYCL.md

@@ -743,6 +743,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 | GGML_SYCL_DISABLE_GRAPH | 0 or 1 (default) | Disable running computations through SYCL Graphs feature. Disabled by default because SYCL Graph is still on development, no better performance. |
 | GGML_SYCL_ENABLE_LEVEL_ZERO | 1 (default) or 0 | Use Level Zero API for device memory allocation instead of SYCL. Reduces system RAM usage on Intel dGPUs by avoiding DMA-buf/TTM host memory staging. Requires GGML_SYCL_SUPPORT_LEVEL_ZERO=ON at build time. |
 | GGML_SYCL_DISABLE_DNN | 0 (default) or 1 | Disable running computations through oneDNN and always use oneMKL. |
+| GGML_SYCL_ENABLE_VMM | 0 or 1 (default) | Enable the virtual-memory device pool. |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |
 | UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS | 0 (default) or 1 | Allow SYCL/Unified Runtime Level Zero device allocations larger than 4 GiB. llama.cpp's direct Level Zero allocation path requests the relaxed maximum-size limit itself when GGML_SYCL_ENABLE_LEVEL_ZERO=1. |
 
@@ -753,6 +754,7 @@ Pass these via `CXXFLAGS` or add a one-off `#define` to enable a flag on the spo
 | Name            | Function                                                                         |
 |-----------------|----------------------------------------------------------------------------------|
 | DEBUG_SYCL_POOL | Enable device memory pool logging on teardown. Useful for profiling allocations. |
+| DEBUG_SYCL_MALLOC | Enable verbose per-call logging of device pool alloc/free operations. |
 
 ## Design Rule
 

docs/backend/ZenDNN.md

@@ -72,10 +72,13 @@ The ZenDNN backend accelerates **matrix multiplication (MUL_MAT)** and **expert-
 |:----------------------:|:-------:|:---------------------------------------------:|
 | FP32                   | Support | Full precision floating point                 |
 | BF16                   | Support | BFloat16 (best performance on Zen 4/Zen 5)    |
+| Q8_0                   | Support | 8-bit quantized weights via [dynamic quantization](https://github.com/amd/ZenDNN/blob/main/docs/operator/lowoha_matmul_operator.md) |
 
 *Notes:*
 
 - **BF16** provides best performance on Zen 4 and Zen 5 EPYC™ processors (Genoa, Turin).
+- **Q8_0** is available for quantized model weights since ZenDNN supports dynamic quantization [LowOHA MatMul operator](https://github.com/amd/ZenDNN/blob/main/docs/operator/lowoha_matmul_operator.md).
+- Other quantization formats fall back to the standard CPU backend unless explicitly supported by the ZenDNN backend.
 
 ## Linux
 
@@ -140,6 +143,15 @@ Download LLaMA 3.1 8B Instruct BF16 model:
 huggingface-cli download meta-llama/Llama-3.1-8B-Instruct-GGUF --local-dir models/
 ```
 
+You can also use a Q8_0 GGUF model:
+
+```sh
+# Download a Q8_0 GGUF model from Hugging Face
+huggingface-cli download meta-llama/Llama-3.1-8B-Instruct-GGUF \
+    Llama-3.1-8B-Instruct-Q8_0.gguf \
+    --local-dir models/
+```
+
 #### 2. Start Server
 
 Run llama.cpp server with ZenDNN acceleration:
@@ -176,6 +188,10 @@ export ZENDNNL_MATMUL_ALGO=1    # Blocked AOCL DLP algo (recommended)
 
 For more details on available algorithms, see the [ZenDNN MatMul Algorithm Documentation](https://github.com/amd/ZenDNN/blob/a18adf8c605fb5f5e52cefd7eda08a7b18febbaf/docs/runtime_env.md#algorithm-details).
 
+### Q8_0 Performance Notes
+
+Q8_0 support is mainly beneficial for prompt processing / prefill workloads where large matrix multiplications dominate execution. Token generation performance may remain close to the standard CPU backend depending on the model, batch size, number of threads, and CPU topology.
+
 ### Profiling and Debugging
 
 For detailed profiling and logging options, refer to the [ZenDNN Logging Documentation](https://github.com/amd/ZenDNN/blob/a18adf8c605fb5f5e52cefd7eda08a7b18febbaf/docs/logging.md).
@@ -184,6 +200,7 @@ For detailed profiling and logging options, refer to the [ZenDNN Logging Documen
 
 - **Limited operation support**: Currently matrix multiplication (MUL_MAT) and expert-based matrix multiplication (MUL_MAT_ID) are accelerated via ZenDNN. Other operations fall back to the standard CPU backend. Future updates may expand supported operations.
 - **BF16 support**: BF16 operations require AMD Zen 4 or Zen 5 architecture (EPYC 9004/9005 series). On older CPUs, operations will use FP32.
+- **Q8_0 support scope**: Q8_0 acceleration is available for supported matrix multiplication paths. Other quantization formats still fall back to the standard CPU backend.
 - **NUMA awareness**: For multi-socket systems, manual NUMA binding may be required for optimal performance.
 
 ## Q&A
@@ -202,7 +219,7 @@ A: ZenDNN is optimized specifically for AMD processors. While it may work on oth
 
 **Q: Does ZenDNN support quantized models?**
 
-A: Currently, ZenDNN primarily supports FP32 and BF16 data types. Quantized model support is not available at this time.
+A: Yes. The ZenDNN backend supports Q8_0 quantized models for supported matrix multiplication operations. FP32 and BF16 are also supported. Other quantization formats may fall back to the standard CPU backend unless explicitly supported by the ZenDNN backend.
 
 **Q: Why is my inference not faster with ZenDNN?**
 

docs/backend/snapdragon/README.md

@@ -10,7 +10,7 @@ This image includes Android NDK, OpenCL SDK, Hexagon SDK, CMake, etc.
 This method works on Linux, macOS, and Windows. macOS and Windows users should install Docker Desktop.
 
 ```
-~/src/llama.cpp$ docker run -it -u $(id -u):$(id -g) --volume $(pwd):/workspace --platform linux/amd64 ghcr.io/snapdragon-toolchain/arm64-android:v0.6
+~/src/llama.cpp$ docker run -it -u $(id -u):$(id -g) --volume $(pwd):/workspace --platform linux/amd64 ghcr.io/snapdragon-toolchain/arm64-android:v0.7
 [d]/> cd /workspace
 ```
 

docs/build.md

@@ -22,6 +22,7 @@ The following sections describe how to build with different backends and options
 * [HIP](#hip)
 * [Vulkan](#vulkan)
 * [CANN](#cann)
+* [ZenDNN](#zendnn)
 * [Arm® KleidiAI™](#arm-kleidiai)
 * [OpenCL](#opencl)
 * [Android](#android-1)

docs/multimodal/granitevision.md

@@ -176,7 +176,7 @@ Note that currently you cannot quantize the visual encoder because granite visio
 
 
 ### 5. Running the Model in Llama cpp
-Build llama cpp normally; you should have a target binary named `llama-mtmd-cli`, which you can pass two binaries to. As an example, we pass the the llama.cpp banner.
+Build llama cpp normally; you should have a target binary named `llama-mtmd-cli`, which you can pass two binaries to. As an example, we pass the llama.cpp banner.
 
 ```bash
 $ ./build/bin/llama-mtmd-cli -m $LLM_GGUF_PATH \

docs/speculative.md

@@ -323,3 +323,8 @@ statistics ngram_map_k: #calls(b,g,a) = 6 1690 26, #gen drafts = 26, #acc drafts
 - `#gen tokens`: number of tokens generated by this implementation (including rejected tokens)
 - `#acc tokens`: number of tokens accepted by the main model
 - `dur(b,g,a): durations of begin (new prompt), generation and accumulation (process acceptance).
+
+## Benchmarking
+
+To measure the end-to-end effect of speculative decoding (throughput, latency, and draft acceptance) across diverse prompts, see the SPEED-Bench client in [tools/server/bench/speed-bench](../tools/server/bench/speed-bench/README.md).
+It runs against a running `llama-server` and can compare a baseline run against a speculative-decoding run.

examples/model-conversion/README.md

@@ -335,7 +335,7 @@ $ make perplexity-run-full QUANTIZED_MODEL=~/path/to/quantized/model-Qxx.gguf LO
 
 ## HuggingFace utilities
 The following targets are useful for creating collections and model repositories
-on Hugging Face in the the ggml-org. These can be used when preparing a release
+on Hugging Face in the ggml-org. These can be used when preparing a release
 to script the process for new model releases.
 
 For the following targets a `HF_TOKEN` environment variable is required.

ggml/CMakeLists.txt

@@ -4,8 +4,8 @@ project("ggml" C CXX ASM)
 
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
-set(GGML_VERSION_MINOR 12)
-set(GGML_VERSION_PATCH 0)
+set(GGML_VERSION_MINOR 13)
+set(GGML_VERSION_PATCH 1)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
 
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")

ggml/include/ggml.h

@@ -1189,8 +1189,8 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
-    // a - x
-    // b - dy
+    // a - dy
+    // b - x
     GGML_API struct ggml_tensor * ggml_silu_back(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,

ggml/src/ggml-alloc.c

@@ -150,7 +150,7 @@ static void ggml_dyn_tallocr_insert_block(struct tallocr_chunk * chunk, size_t o
 
 static void ggml_dyn_tallocr_remove_block(struct tallocr_chunk * chunk, int idx) {
     // shift all elements after idx by 1 to the left, overwriting the element at idx
-    for (int i = idx; i < chunk->n_free_blocks; i++) {
+    for (int i = idx; i < chunk->n_free_blocks - 1; i++) {
         chunk->free_blocks[i] = chunk->free_blocks[i+1];
     }
     chunk->n_free_blocks--;

ggml/src/ggml-backend-meta.cpp

@@ -2076,6 +2076,7 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
             node_zero->src[0] = node;
             ggml_set_op_params_f32(node_zero, 0, 0.0f);
             node_zero->data = node->data;
+            node_zero->buffer = node->buffer;
             node_zero->flags |= GGML_TENSOR_FLAG_COMPUTE;
 
             step_cgraphs[j] = get_cgraph_aux();

ggml/src/ggml-cpu/arch/loongarch/quants.c

@@ -977,6 +977,35 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     sumf = hsum_float_8(acc);
 
     *s = sumf;
+
+#elif defined(__loongarch_sx)
+
+    __m128 acc = (__m128)__lsx_vldi(0);
+
+    for (; ib < nb; ++ib) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d);
+        const __m128i qx_0 = __lsx_vld((const __m128i *)x[ib].qs, 0);
+        const __m128i qx_1 = __lsx_vld((const __m128i *)x[ib].qs + 1, 0);
+        const __m128i qy_0 = __lsx_vld((const __m128i *)y[ib].qs, 0);
+        const __m128i qy_1 = __lsx_vld((const __m128i *)y[ib].qs + 1, 0);
+
+        const __m128i p16_0 = lsx_maddubs_h(qx_0, qy_0);
+        const __m128i p16_1 = lsx_maddubs_h(qx_1, qy_1);
+
+        // Sum int16 pairs → int32
+        const __m128i s_0 = __lsx_vaddwev_w_h(p16_0, p16_1);
+        const __m128i s_1 = __lsx_vaddwod_w_h(p16_0, p16_1);
+
+        const __m128 q = __lsx_vffint_s_w(__lsx_vadd_w(s_0, s_1));
+        acc = __lsx_vfmadd_s(__lsx_vreplfr2vr_s(d), q, acc);
+    }
+
+    __m128 res = lsx_hadd_s(acc, acc);
+    res = lsx_hadd_s(res, res);
+    sumf = ((v4f32)res)[0];
+
+    *s = sumf;
+
 #else
     UNUSED(nb);
     UNUSED(ib);
@@ -1443,6 +1472,99 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     *s = hsum_float_8(acc);
 
+#elif defined(__loongarch_sx)
+
+    const __m128i m32s = __lsx_vreplgr2vr_b(32);
+
+    __m128 acc_0 = (__m128)__lsx_vldi(0);
+    __m128 acc_1 = (__m128)__lsx_vldi(0);
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
+
+        const __m128i scale_i8 = __lsx_vld(x[i].scales, 0);
+        const __m128i scales_lo = __lsx_vsllwil_h_b(scale_i8, 0);
+        const __m128i scales_hi = __lsx_vsllwil_h_b(__lsx_vbsrl_v(scale_i8, 8), 0);
+
+        __m128i sumi_0 = __lsx_vldi(0);
+        __m128i sumi_1 = __lsx_vldi(0);
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            const __m128i q4bitsH_0 = __lsx_vld((const __m128i*)qh, 0); qh += 16;
+            const __m128i q4bitsH_1 = __lsx_vld((const __m128i*)qh, 0); qh += 16;
+
+            const __m128i q4h_0 = __lsx_vslli_b(__lsx_vandi_b(q4bitsH_0, 3), 4);
+            const __m128i q4h_1 = __lsx_vslli_b(__lsx_vandi_b(q4bitsH_1, 3), 4);
+            const __m128i q4h_2 = __lsx_vslli_b(__lsx_vandi_b(q4bitsH_0, 3 << 2), 2);
+            const __m128i q4h_3 = __lsx_vslli_b(__lsx_vandi_b(q4bitsH_1, 3 << 2), 2);
+            const __m128i q4h_4 = __lsx_vandi_b(q4bitsH_0, 3 << 4);
+            const __m128i q4h_5 = __lsx_vandi_b(q4bitsH_1, 3 << 4);
+            const __m128i q4h_6 = __lsx_vsrli_b(__lsx_vandi_b(q4bitsH_0, 3 << 6), 2);
+            const __m128i q4h_7 = __lsx_vsrli_b(__lsx_vandi_b(q4bitsH_1, 3 << 6), 2);
+
+            const __m128i q4bits1_0 = __lsx_vld((const __m128i*)q4, 0); q4 += 16;
+            const __m128i q4bits1_1 = __lsx_vld((const __m128i*)q4, 0); q4 += 16;
+            const __m128i q4bits2_0 = __lsx_vld((const __m128i*)q4, 0); q4 += 16;
+            const __m128i q4bits2_1 = __lsx_vld((const __m128i*)q4, 0); q4 += 16;
+
+            const __m128i q4_0 = __lsx_vor_v(__lsx_vandi_b(q4bits1_0, 0xf), q4h_0);
+            const __m128i q4_1 = __lsx_vor_v(__lsx_vandi_b(q4bits1_1, 0xf), q4h_1);
+            const __m128i q4_2 = __lsx_vor_v(__lsx_vandi_b(q4bits2_0, 0xf), q4h_2);
+            const __m128i q4_3 = __lsx_vor_v(__lsx_vandi_b(q4bits2_1, 0xf), q4h_3);
+            const __m128i q4_4 = __lsx_vor_v(__lsx_vsrli_b(q4bits1_0, 4), q4h_4);
+            const __m128i q4_5 = __lsx_vor_v(__lsx_vsrli_b(q4bits1_1, 4), q4h_5);
+            const __m128i q4_6 = __lsx_vor_v(__lsx_vsrli_b(q4bits2_0, 4), q4h_6);
+            const __m128i q4_7 = __lsx_vor_v(__lsx_vsrli_b(q4bits2_1, 4), q4h_7);
+
+            const __m128i q8_0 = __lsx_vld((const __m128i*)q8, 0); q8 += 16;
+            const __m128i q8_1 = __lsx_vld((const __m128i*)q8, 0); q8 += 16;
+            const __m128i q8_2 = __lsx_vld((const __m128i*)q8, 0); q8 += 16;
+            const __m128i q8_3 = __lsx_vld((const __m128i*)q8, 0); q8 += 16;
+            const __m128i q8_4 = __lsx_vld((const __m128i*)q8, 0); q8 += 16;
+            const __m128i q8_5 = __lsx_vld((const __m128i*)q8, 0); q8 += 16;
+            const __m128i q8_6 = __lsx_vld((const __m128i*)q8, 0); q8 += 16;
+            const __m128i q8_7 = __lsx_vld((const __m128i*)q8, 0); q8 += 16;
+
+            __m128i p16_0 = lsx_maddubs_h(__lsx_vsub_b(q4_0, m32s), q8_0);
+            __m128i p16_1 = lsx_maddubs_h(__lsx_vsub_b(q4_1, m32s), q8_1);
+            __m128i p16_2 = lsx_maddubs_h(__lsx_vsub_b(q4_2, m32s), q8_2);
+            __m128i p16_3 = lsx_maddubs_h(__lsx_vsub_b(q4_3, m32s), q8_3);
+            __m128i p16_4 = lsx_maddubs_h(__lsx_vsub_b(q4_4, m32s), q8_4);
+            __m128i p16_5 = lsx_maddubs_h(__lsx_vsub_b(q4_5, m32s), q8_5);
+            __m128i p16_6 = lsx_maddubs_h(__lsx_vsub_b(q4_6, m32s), q8_6);
+            __m128i p16_7 = lsx_maddubs_h(__lsx_vsub_b(q4_7, m32s), q8_7);
+
+            const __m128i sc_vec = j == 0 ? scales_lo : scales_hi;
+
+            p16_0 = lsx_madd_h(__lsx_vreplvei_h(sc_vec, 0), p16_0);
+            p16_1 = lsx_madd_h(__lsx_vreplvei_h(sc_vec, 1), p16_1);
+            p16_2 = lsx_madd_h(__lsx_vreplvei_h(sc_vec, 2), p16_2);
+            p16_3 = lsx_madd_h(__lsx_vreplvei_h(sc_vec, 3), p16_3);
+            p16_4 = lsx_madd_h(__lsx_vreplvei_h(sc_vec, 4), p16_4);
+            p16_5 = lsx_madd_h(__lsx_vreplvei_h(sc_vec, 5), p16_5);
+            p16_6 = lsx_madd_h(__lsx_vreplvei_h(sc_vec, 6), p16_6);
+            p16_7 = lsx_madd_h(__lsx_vreplvei_h(sc_vec, 7), p16_7);
+
+            sumi_0 = __lsx_vadd_w(sumi_0, __lsx_vadd_w(p16_0, p16_2));
+            sumi_1 = __lsx_vadd_w(sumi_1, __lsx_vadd_w(p16_1, p16_3));
+            sumi_0 = __lsx_vadd_w(sumi_0, __lsx_vadd_w(p16_4, p16_6));
+            sumi_1 = __lsx_vadd_w(sumi_1, __lsx_vadd_w(p16_5, p16_7));
+        }
+
+        __m128 p_0 = __lsx_vfmul_s(__lsx_vreplfr2vr_s(d), __lsx_vffint_s_w(sumi_0));
+        __m128 p_1 = __lsx_vfmul_s(__lsx_vreplfr2vr_s(d), __lsx_vffint_s_w(sumi_1));
+        acc_0 = __lsx_vfadd_s(p_0, acc_0);
+        acc_1 = __lsx_vfadd_s(p_1, acc_1);
+    }
+
+    *s = hsum_float_4x4(acc_0, acc_1, (__m128)__lsx_vldi(0), (__m128)__lsx_vldi(0));
+
 #else
     UNUSED(x);
     UNUSED(y);
@@ -2149,6 +2271,35 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
     *s = hsum_float_8(accum);
 
+#elif defined(__loongarch_sx)
+
+    const __m128i values128 = __lsx_vld((const __m128i*)kvalues_iq4nl, 0);
+
+    __m128 accum = (__m128)__lsx_vldi(0);
+    for (int ibl = 0; ibl < nb; ++ibl) {
+        const uint8_t * qs = x[ibl].qs;
+        const int8_t  * q8 = y[ibl].qs;
+        uint16_t sh = x[ibl].scales_h;
+        __m128i sumi = __lsx_vldi(0);
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            const __m128i q4bits = __lsx_vld((const __m128i*)qs, 0); qs += 16;
+            const __m128i q8b_0 = __lsx_vld((const __m128i*)q8, 0); q8 += 16;
+            const __m128i q8b_1 = __lsx_vld((const __m128i*)q8, 0); q8 += 16;
+            const __m128i q4b_0 = __lsx_vshuf_b(values128, values128, __lsx_vandi_b(q4bits, 0xf));
+            const __m128i q4b_1 = __lsx_vshuf_b(values128, values128, __lsx_vsrli_b(q4bits, 4));
+            const __m128i p16_0 = lsx_maddubs_h(q4b_0, q8b_0);
+            const __m128i p16_1 = lsx_maddubs_h(q4b_1, q8b_1);
+            const int16_t ls = (((x[ibl].scales_l[ib/2] >> ((ib & 1) * 4)) & 0xf) | ((sh & 0x3) << 4)) - 32;
+            sh >>= 2;
+            sumi = __lsx_vadd_w(lsx_madd_h(p16_0, __lsx_vreplgr2vr_h(ls)), sumi);
+            sumi = __lsx_vadd_w(lsx_madd_h(p16_1, __lsx_vreplgr2vr_h(ls)), sumi);
+        }
+        const float ds = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        accum = __lsx_vfadd_s(__lsx_vfmul_s(__lsx_vreplfr2vr_s(ds), __lsx_vffint_s_w(sumi)), accum);
+    }
+
+    *s = ((v4f32)lsx_hadd_s(lsx_hadd_s(accum, accum), lsx_hadd_s(accum, accum)))[0];
+
 #else
     UNUSED(x);
     UNUSED(y);

ggml/src/ggml-cpu/ops.cpp

@@ -2235,8 +2235,42 @@ static void ggml_compute_forward_fill_f32(const ggml_compute_params * params, gg
     }
 }
 
+static void ggml_compute_forward_fill_f16(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_fp16_t c = GGML_CPU_FP32_TO_FP16(ggml_get_op_params_f32(dst, 0));
+
+    GGML_TENSOR_LOCALS(int64_t, ne, dst, ne);
+    GGML_TENSOR_LOCALS(size_t,  nb, dst, nb);
+
+    const auto [ir0, ir1] = get_thread_range(params, dst);
+
+    for (int64_t ir = ir0; ir < ir1; ++ir) {
+        const int64_t i03 = ir/(ne2*ne1);
+        const int64_t i02 = (ir - i03*ne2*ne1)/ne1;
+        const int64_t i01 = (ir - i03*ne2*ne1 - i02*ne1);
+
+        ggml_fp16_t * dst_ptr  = (ggml_fp16_t *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1);
+
+        ggml_vec_set_f16(ne0, dst_ptr, c);
+    }
+}
+
 void ggml_compute_forward_fill(const ggml_compute_params * params, ggml_tensor * dst) {
-    ggml_compute_forward_fill_f32(params, dst);
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_fill_f32(params, dst);
+            } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_fill_f16(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("unsupported type for ggml_compute_forward_fill: %s", ggml_type_name(src0->type));
+            }
+    }
 }
 
 // ggml_compute_tri

ggml/src/ggml-cpu/simd-mappings.h

@@ -1125,25 +1125,12 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) {
 #define GGML_F16_EPR  4
 
 static inline __m128 __lsx_f16x4_load(const ggml_fp16_t * x) {
-    float tmp[4];
-
-    tmp[0] = GGML_CPU_FP16_TO_FP32(x[0]);
-    tmp[1] = GGML_CPU_FP16_TO_FP32(x[1]);
-    tmp[2] = GGML_CPU_FP16_TO_FP32(x[2]);
-    tmp[3] = GGML_CPU_FP16_TO_FP32(x[3]);
-
-    return (__m128)__lsx_vld(tmp, 0);
+    return __lsx_vfcvtl_s_h(__lsx_vld((const void *)x, 0));
 }
 
 static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
-    float arr[4];
-
-    __lsx_vst(y, arr, 0);
-
-    x[0] = GGML_CPU_FP32_TO_FP16(arr[0]);
-    x[1] = GGML_CPU_FP32_TO_FP16(arr[1]);
-    x[2] = GGML_CPU_FP32_TO_FP16(arr[2]);
-    x[3] = GGML_CPU_FP32_TO_FP16(arr[3]);
+    __m128i a = __lsx_vfcvt_h_s(y, y);
+    memcpy(x, &a, sizeof(ggml_fp16_t) * 4);
 }
 
 #define GGML_F32Cx4             __m128

ggml/src/ggml-cpu/vec.cpp

@@ -273,67 +273,51 @@ void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * G
 
 #if defined(GGML_SIMD)
     #if defined(__ARM_FEATURE_SVE)
-        const int sve_register_length = svcntb() * 8; //get vector length
-        const int ggml_f16_epr = sve_register_length / 16; // running when 16
-        const int ggml_f16_step = 8 * ggml_f16_epr; // choose 8 SVE registers
+        const int ggml_f16_epr = svcnth();
+        const int ggml_f16_step = 8 * ggml_f16_epr;
+        const int np = n - (n % ggml_f16_step);
+        const int np2 = n - (n % ggml_f16_epr);
 
-        const int np= (n & ~(ggml_f16_step - 1));
-        svfloat16_t sum1 = svdup_n_f16(0.0f);
-        svfloat16_t sum2 = svdup_n_f16(0.0f);
-        svfloat16_t sum3 = svdup_n_f16(0.0f);
-        svfloat16_t sum4 = svdup_n_f16(0.0f);
+        svfloat32_t sum1_lo = svdup_n_f32(0.0f);
+        svfloat32_t sum1_hi = svdup_n_f32(0.0f);
+        svfloat32_t sum2_lo = svdup_n_f32(0.0f);
+        svfloat32_t sum2_hi = svdup_n_f32(0.0f);
+        svfloat32_t sum3_lo = svdup_n_f32(0.0f);
+        svfloat32_t sum3_hi = svdup_n_f32(0.0f);
+        svfloat32_t sum4_lo = svdup_n_f32(0.0f);
+        svfloat32_t sum4_hi = svdup_n_f32(0.0f);
 
-        svfloat16_t ax1, ax2, ax3, ax4, ax5, ax6, ax7, ax8;
-        svfloat16_t ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8;
         for (int i = 0; i < np; i += ggml_f16_step) {
-            ax1 = GGML_F16x_VEC_LOAD(x + i + 0 * ggml_f16_epr, 0);
-            ay1 = GGML_F16x_VEC_LOAD(y + i + 0 * ggml_f16_epr, 0);
-            sum1 = GGML_F16x_VEC_FMA(sum1, ax1, ay1);
-
-            ax2 = GGML_F16x_VEC_LOAD(x + i + 1 * ggml_f16_epr, 1);
-            ay2 = GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 1);
-            sum2 = GGML_F16x_VEC_FMA(sum2, ax2, ay2);
-
-            ax3 = GGML_F16x_VEC_LOAD(x + i + 2 * ggml_f16_epr, 2);
-            ay3 = GGML_F16x_VEC_LOAD(y + i + 2 * ggml_f16_epr, 2);
-            sum3 = GGML_F16x_VEC_FMA(sum3, ax3, ay3);
-
-            ax4 = GGML_F16x_VEC_LOAD(x + i + 3 * ggml_f16_epr, 3);
-            ay4 = GGML_F16x_VEC_LOAD(y + i + 3 * ggml_f16_epr, 3);
-            sum4 = GGML_F16x_VEC_FMA(sum4, ax4, ay4);
-
-            ax5 = GGML_F16x_VEC_LOAD(x + i + 4 * ggml_f16_epr, 4);
-            ay5 = GGML_F16x_VEC_LOAD(y + i + 4 * ggml_f16_epr, 4);
-            sum1 = GGML_F16x_VEC_FMA(sum1, ax5, ay5);
-
-            ax6 = GGML_F16x_VEC_LOAD(x + i + 5 * ggml_f16_epr, 5);
-            ay6 = GGML_F16x_VEC_LOAD(y + i + 5 * ggml_f16_epr, 5);
-            sum2 = GGML_F16x_VEC_FMA(sum2, ax6, ay6);
-
-            ax7 = GGML_F16x_VEC_LOAD(x + i + 6 * ggml_f16_epr, 6);
-            ay7 = GGML_F16x_VEC_LOAD(y + i + 6 * ggml_f16_epr, 6);
-            sum3 = GGML_F16x_VEC_FMA(sum3, ax7, ay7);
-
-            ax8 = GGML_F16x_VEC_LOAD(x + i + 7 * ggml_f16_epr, 7);
-            ay8 = GGML_F16x_VEC_LOAD(y + i + 7 * ggml_f16_epr, 7);
-            sum4 = GGML_F16x_VEC_FMA(sum4, ax8, ay8);
+            ggml_sve_f16_fma_widened(&sum1_lo, &sum1_hi, GGML_F16x_VEC_LOAD(x + i + 0 * ggml_f16_epr, 0), GGML_F16x_VEC_LOAD(y + i + 0 * ggml_f16_epr, 0));
+            ggml_sve_f16_fma_widened(&sum2_lo, &sum2_hi, GGML_F16x_VEC_LOAD(x + i + 1 * ggml_f16_epr, 1), GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 1));
+            ggml_sve_f16_fma_widened(&sum3_lo, &sum3_hi, GGML_F16x_VEC_LOAD(x + i + 2 * ggml_f16_epr, 2), GGML_F16x_VEC_LOAD(y + i + 2 * ggml_f16_epr, 2));
+            ggml_sve_f16_fma_widened(&sum4_lo, &sum4_hi, GGML_F16x_VEC_LOAD(x + i + 3 * ggml_f16_epr, 3), GGML_F16x_VEC_LOAD(y + i + 3 * ggml_f16_epr, 3));
+            ggml_sve_f16_fma_widened(&sum1_lo, &sum1_hi, GGML_F16x_VEC_LOAD(x + i + 4 * ggml_f16_epr, 4), GGML_F16x_VEC_LOAD(y + i + 4 * ggml_f16_epr, 4));
+            ggml_sve_f16_fma_widened(&sum2_lo, &sum2_hi, GGML_F16x_VEC_LOAD(x + i + 5 * ggml_f16_epr, 5), GGML_F16x_VEC_LOAD(y + i + 5 * ggml_f16_epr, 5));
+            ggml_sve_f16_fma_widened(&sum3_lo, &sum3_hi, GGML_F16x_VEC_LOAD(x + i + 6 * ggml_f16_epr, 6), GGML_F16x_VEC_LOAD(y + i + 6 * ggml_f16_epr, 6));
+            ggml_sve_f16_fma_widened(&sum4_lo, &sum4_hi, GGML_F16x_VEC_LOAD(x + i + 7 * ggml_f16_epr, 7), GGML_F16x_VEC_LOAD(y + i + 7 * ggml_f16_epr, 7));
         }
 
-        const int np2 = (n & ~(ggml_f16_epr - 1)); // round down to multiple of 8
-        for (int k = np; k < np2; k += ggml_f16_epr) {
-            svfloat16_t rx = GGML_F16x_VEC_LOAD(x + k, 0);
-            svfloat16_t ry = GGML_F16x_VEC_LOAD(y + k, 0);
-            sum1 = GGML_F16x_VEC_FMA(sum1, rx, ry);
+        for (int i = np; i < np2; i += ggml_f16_epr) {
+            ggml_sve_f16_fma_widened(&sum1_lo, &sum1_hi, GGML_F16x_VEC_LOAD(x + i, 0), GGML_F16x_VEC_LOAD(y + i, 0));
         }
 
         if (np2 < n) {
-            svbool_t pg = svwhilelt_b16(np2, n);
-            svfloat16_t hx = svld1_f16(pg, (const __fp16 *)(x + np2));
-            svfloat16_t hy = svld1_f16(pg, (const __fp16 *)(y + np2));
+            const svbool_t pg = svwhilelt_b16(np2, n);
+            const svfloat16_t rx = svld1_f16(pg, (const __fp16 *)(x + np2));
+            const svfloat16_t ry = svld1_f16(pg, (const __fp16 *)(y + np2));
 
-            sum1 = svmad_f16_x(pg, hx, hy, sum1);
+            ggml_sve_f16_fma_widened(&sum1_lo, &sum1_hi, rx, ry);
         }
-        GGML_F16x_VEC_REDUCE(sumf, sum1, sum2, sum3, sum4);
+
+        sum1_lo = svadd_f32_m(DEFAULT_PG32, sum1_lo, sum2_lo);
+        sum1_hi = svadd_f32_m(DEFAULT_PG32, sum1_hi, sum2_hi);
+        sum3_lo = svadd_f32_m(DEFAULT_PG32, sum3_lo, sum4_lo);
+        sum3_hi = svadd_f32_m(DEFAULT_PG32, sum3_hi, sum4_hi);
+        sum1_lo = svadd_f32_m(DEFAULT_PG32, sum1_lo, sum3_lo);
+        sum1_hi = svadd_f32_m(DEFAULT_PG32, sum1_hi, sum3_hi);
+
+        sumf = ggml_sve_sum_f32x2(sum1_lo, sum1_hi);
     #elif defined(__riscv_v_intrinsic)
         #if defined(__riscv_zvfh)
             int vl = __riscv_vsetvlmax_e32m2();

ggml/src/ggml-cpu/vec.h

@@ -14,6 +14,35 @@
 // floating point type used to accumulate sums
 typedef double ggml_float;
 
+#if defined(__ARM_FEATURE_SVE)
+inline static void ggml_sve_f16_fma_widened(
+        svfloat32_t * acc_lo,
+        svfloat32_t * acc_hi,
+        svfloat16_t x,
+        svfloat16_t y) {
+#if defined(__ARM_FEATURE_SVE2)
+    *acc_lo = svmlalb_f32(*acc_lo, x, y);
+    *acc_hi = svmlalt_f32(*acc_hi, x, y);
+#else
+    // Plain SVE fallback path if SVE2 instructions not available
+    svfloat16_t x_even = svtrn1_f16(x, x);
+    svfloat16_t x_odd = svtrn2_f16(x, x);
+
+    svfloat16_t y_even = svtrn1_f16(y, y);
+    svfloat16_t y_odd = svtrn2_f16(y, y);
+
+    svbool_t pg = svptrue_b32();
+
+    *acc_lo = svmla_f32_x(pg, *acc_lo, svcvt_f32_f16_x(pg, x_even), svcvt_f32_f16_x(pg, y_even));
+    *acc_hi = svmla_f32_x(pg, *acc_hi, svcvt_f32_f16_x(pg, x_odd), svcvt_f32_f16_x(pg, y_odd));
+#endif
+}
+
+inline static ggml_float ggml_sve_sum_f32x2(svfloat32_t sum_lo, svfloat32_t sum_hi) {
+    return (ggml_float) (svaddv_f32(svptrue_b32(), sum_lo) + svaddv_f32(svptrue_b32(), sum_hi));
+}
+#endif
+
 #define GGML_GELU_FP16
 #define GGML_GELU_QUICK_FP16
 
@@ -122,108 +151,61 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
 #if defined(GGML_SIMD)
     #if defined(__ARM_FEATURE_SVE)
 
-        const int sve_register_length = svcntb() * 8;
-        const int ggml_f16_epr = sve_register_length / 16; // running when 16
-        const int ggml_f16_step = 8 * ggml_f16_epr; // choose 8 SVE registers
+        const int ggml_f16_epr = svcnth();
+        const int ggml_f16_step = 2 * ggml_f16_epr;
+        int np = n - (n % ggml_f16_step);
+        int np2 = n - (n % ggml_f16_epr);
 
-        int np = (n & ~(ggml_f16_step - 1));
-
-        svfloat16_t sum_00 = svdup_n_f16(0.0f);
-        svfloat16_t sum_01 = svdup_n_f16(0.0f);
-        svfloat16_t sum_02 = svdup_n_f16(0.0f);
-        svfloat16_t sum_03 = svdup_n_f16(0.0f);
-
-        svfloat16_t sum_10 = svdup_n_f16(0.0f);
-        svfloat16_t sum_11 = svdup_n_f16(0.0f);
-        svfloat16_t sum_12 = svdup_n_f16(0.0f);
-        svfloat16_t sum_13 = svdup_n_f16(0.0f);
-
-        svfloat16_t ax1, ax2, ax3, ax4, ax5, ax6, ax7, ax8;
-        svfloat16_t ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8;
+        svfloat32_t sum_0_0_lo = svdup_n_f32(0.0f);
+        svfloat32_t sum_0_0_hi = svdup_n_f32(0.0f);
+        svfloat32_t sum_0_1_lo = svdup_n_f32(0.0f);
+        svfloat32_t sum_0_1_hi = svdup_n_f32(0.0f);
+        svfloat32_t sum_1_0_lo = svdup_n_f32(0.0f);
+        svfloat32_t sum_1_0_hi = svdup_n_f32(0.0f);
+        svfloat32_t sum_1_1_lo = svdup_n_f32(0.0f);
+        svfloat32_t sum_1_1_hi = svdup_n_f32(0.0f);
 
         for (int i = 0; i < np; i += ggml_f16_step) {
-            ay1 = GGML_F16x_VEC_LOAD(y + i + 0 * ggml_f16_epr, 0); // 8 elements
+            const svfloat16_t ay0 = GGML_F16x_VEC_LOAD(y + i, 0);
+            const svfloat16_t ax00 = GGML_F16x_VEC_LOAD(x[0] + i, 0);
+            const svfloat16_t ax01 = GGML_F16x_VEC_LOAD(x[1] + i, 0);
 
-            ax1 = GGML_F16x_VEC_LOAD(x[0] + i + 0*ggml_f16_epr, 0); // 8 elements
-            sum_00 = GGML_F16x_VEC_FMA(sum_00, ax1, ay1);     // sum_00 = sum_00+ax1*ay1
-            ax1 = GGML_F16x_VEC_LOAD(x[1] + i + 0*ggml_f16_epr, 0); // 8 elements
-            sum_10 = GGML_F16x_VEC_FMA(sum_10, ax1, ay1);
+            ggml_sve_f16_fma_widened(&sum_0_0_lo, &sum_0_0_hi, ax00, ay0);
+            ggml_sve_f16_fma_widened(&sum_1_0_lo, &sum_1_0_hi, ax01, ay0);
 
-            ay2 = GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 1); // next 8 elements
+            const svfloat16_t ay1 = GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 0);
+            const svfloat16_t ax10 = GGML_F16x_VEC_LOAD(x[0] + i + 1 * ggml_f16_epr, 0);
+            const svfloat16_t ax11 = GGML_F16x_VEC_LOAD(x[1] + i + 1 * ggml_f16_epr, 0);
 
-            ax2 = GGML_F16x_VEC_LOAD(x[0] + i + 1*ggml_f16_epr, 1); // next 8 elements
-            sum_01 = GGML_F16x_VEC_FMA(sum_01, ax2, ay2);
-            ax2 = GGML_F16x_VEC_LOAD(x[1] + i + 1*ggml_f16_epr, 1);
-            sum_11 = GGML_F16x_VEC_FMA(sum_11, ax2, ay2);
-
-            ay3 = GGML_F16x_VEC_LOAD(y + i + 2 * ggml_f16_epr, 2);
-
-            ax3 = GGML_F16x_VEC_LOAD(x[0] + i + 2*ggml_f16_epr, 2);
-            sum_02 = GGML_F16x_VEC_FMA(sum_02, ax3, ay3);
-            ax3 = GGML_F16x_VEC_LOAD(x[1] + i + 2*ggml_f16_epr, 2);
-            sum_12 = GGML_F16x_VEC_FMA(sum_12, ax3, ay3);
-
-            ay4 = GGML_F16x_VEC_LOAD(y + i + 3 * ggml_f16_epr, 3);
-
-            ax4 = GGML_F16x_VEC_LOAD(x[0] + i + 3*ggml_f16_epr, 3);
-            sum_03 = GGML_F16x_VEC_FMA(sum_03, ax4, ay4);
-            ax4 = GGML_F16x_VEC_LOAD(x[1] + i + 3*ggml_f16_epr, 3);
-            sum_13 = GGML_F16x_VEC_FMA(sum_13, ax4, ay4);
-
-            ay5 = GGML_F16x_VEC_LOAD(y + i + 4 * ggml_f16_epr, 4);
-
-            ax5 = GGML_F16x_VEC_LOAD(x[0] + i + 4*ggml_f16_epr, 4);
-
-            sum_00 = GGML_F16x_VEC_FMA(sum_00, ax5, ay5);
-            ax5 = GGML_F16x_VEC_LOAD(x[1] + i + 4*ggml_f16_epr, 4);
-            sum_10 = GGML_F16x_VEC_FMA(sum_10, ax5, ay5);
-
-            ay6 = GGML_F16x_VEC_LOAD(y + i + 5 * ggml_f16_epr, 5);
-
-            ax6 = GGML_F16x_VEC_LOAD(x[0] + i + 5*ggml_f16_epr, 5);
-
-            sum_01 = GGML_F16x_VEC_FMA(sum_01, ax6, ay6);
-            ax6 = GGML_F16x_VEC_LOAD(x[1] + i + 5*ggml_f16_epr, 5);
-            sum_11 = GGML_F16x_VEC_FMA(sum_11, ax6, ay6);
-
-            ay7 = GGML_F16x_VEC_LOAD(y + i + 6 * ggml_f16_epr, 6);
-
-            ax7 = GGML_F16x_VEC_LOAD(x[0] + i + 6*ggml_f16_epr, 6);
-
-            sum_02 = GGML_F16x_VEC_FMA(sum_02, ax7, ay7);
-            ax7 = GGML_F16x_VEC_LOAD(x[1] + i + 6*ggml_f16_epr, 6);
-            sum_12 = GGML_F16x_VEC_FMA(sum_12, ax7, ay7);
-
-            ay8 = GGML_F16x_VEC_LOAD(y + i + 7 * ggml_f16_epr, 7);
-
-            ax8 = GGML_F16x_VEC_LOAD(x[0] + i + 7*ggml_f16_epr, 7);
-
-            sum_03 = GGML_F16x_VEC_FMA(sum_03, ax8, ay8);
-            ax8 = GGML_F16x_VEC_LOAD(x[1] + i + 7*ggml_f16_epr, 7);
-            sum_13 = GGML_F16x_VEC_FMA(sum_13, ax8, ay8);
+            ggml_sve_f16_fma_widened(&sum_0_1_lo, &sum_0_1_hi, ax10, ay1);
+            ggml_sve_f16_fma_widened(&sum_1_1_lo, &sum_1_1_hi, ax11, ay1);
         }
 
-        const int np2 = (n & ~(ggml_f16_epr - 1));
-        for (int k = np; k < np2; k += ggml_f16_epr) {
-            svfloat16_t ry = GGML_F16x_VEC_LOAD(y + k, 0);
+        for (int i = np; i < np2; i += ggml_f16_epr) {
+            const svfloat16_t ry = GGML_F16x_VEC_LOAD(y + i, 0);
+            const svfloat16_t rx0 = GGML_F16x_VEC_LOAD(x[0] + i, 0);
+            const svfloat16_t rx1 = GGML_F16x_VEC_LOAD(x[1] + i, 0);
 
-            svfloat16_t rx = GGML_F16x_VEC_LOAD(x[0] + k, 0);
-            sum_00 = GGML_F16x_VEC_FMA(sum_00, rx, ry);
-            rx = GGML_F16x_VEC_LOAD(x[1] + k, 0);
-            sum_10 = GGML_F16x_VEC_FMA(sum_10, rx, ry);
+            ggml_sve_f16_fma_widened(&sum_0_0_lo, &sum_0_0_hi, rx0, ry);
+            ggml_sve_f16_fma_widened(&sum_1_0_lo, &sum_1_0_hi, rx1, ry);
         }
 
         if (np2 < n) {
-            svbool_t pg = svwhilelt_b16(np2, n);
-            svfloat16_t hx_0 = svld1_f16(pg, (const __fp16 *)(x[0] + np2));
-            svfloat16_t hx_1 = svld1_f16(pg, (const __fp16 *)(x[1] + np2));
-            svfloat16_t hy = svld1_f16(pg, (const __fp16 *)(y + np2));
+            const svbool_t pg = svwhilelt_b16(np2, n);
+            const svfloat16_t ay = svld1_f16(pg, (const __fp16 *)(y + np2));
+            const svfloat16_t ax0 = svld1_f16(pg, (const __fp16 *)(x[0] + np2));
+            const svfloat16_t ax1 = svld1_f16(pg, (const __fp16 *)(x[1] + np2));
 
-            sum_00 = svmad_f16_x(pg, hx_0, hy, sum_00);
-            sum_10 = svmad_f16_x(pg, hx_1, hy, sum_10);
+            ggml_sve_f16_fma_widened(&sum_0_0_lo, &sum_0_0_hi, ax0, ay);
+            ggml_sve_f16_fma_widened(&sum_1_0_lo, &sum_1_0_hi, ax1, ay);
         }
-        GGML_F16x_VEC_REDUCE(sumf[0], sum_00, sum_01, sum_02, sum_03);
-        GGML_F16x_VEC_REDUCE(sumf[1], sum_10, sum_11, sum_12, sum_13);
+
+        svfloat32_t sum_0_lo = svadd_f32_x(DEFAULT_PG32, sum_0_0_lo, sum_0_1_lo);
+        svfloat32_t sum_0_hi = svadd_f32_x(DEFAULT_PG32, sum_0_0_hi, sum_0_1_hi);
+        svfloat32_t sum_1_lo = svadd_f32_x(DEFAULT_PG32, sum_1_0_lo, sum_1_1_lo);
+        svfloat32_t sum_1_hi = svadd_f32_x(DEFAULT_PG32, sum_1_0_hi, sum_1_1_hi);
+        sumf[0] = ggml_sve_sum_f32x2(sum_0_lo, sum_0_hi);
+        sumf[1] = ggml_sve_sum_f32x2(sum_1_lo, sum_1_hi);
         np = n;
     #elif defined(__riscv_v_intrinsic)
         #if defined(__riscv_zvfh)

ggml/src/ggml-cuda/common.cuh

@@ -7,6 +7,7 @@
 #include <cstdint>
 #include <cstdlib>
 #include <memory>
+#include <mutex>
 
 #if defined(GGML_USE_HIP)
 #define GGML_COMMON_DECL_HIP
@@ -110,11 +111,14 @@
 #    define GGML_CUDA_USE_CUB
 #endif  // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
 
-// PDL host-side support (cudaLaunchKernelEx) requires CUDART >= 11.8 and excludes HIP/MUSA.
+// PDL host-side support (cudaLaunchKernelEx) requires CUDART >= 11.8.
+// However, this has been bugged in CTK < 12.3 for MSVC builds, see
+// https://github.com/ggml-org/llama.cpp/pull/22522#discussion_r3302393293
 // __CUDA_ARCH__  is undefined in host passes; GPU arch check happens in device-side code.
-#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11080
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && \
+    (CUDART_VERSION >= 12030 || (!(defined(_MSC_VER) && !defined(__clang__)) && CUDART_VERSION >= 11080))
 #    define GGML_CUDA_USE_PDL
-#endif  // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11080
+#endif  // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && (CUDART_VERSION >= 12030 || (!(defined(_MSC_VER) && !defined(__clang__)) && CUDART_VERSION >= 11080))
 
 static __device__ __forceinline__ void ggml_cuda_pdl_sync() {
 #if defined(GGML_CUDA_USE_PDL) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= GGML_CUDA_CC_HOPPER
@@ -1549,6 +1553,62 @@ struct ggml_cuda_pdl_config {
     ggml_cuda_pdl_config& operator=(ggml_cuda_pdl_config&&) = delete;
 
 };
+
+static bool ggml_cuda_kernel_can_use_pdl(const void * kernel) {
+    const int device = ggml_cuda_get_device();
+
+    struct cache_key {
+        int          device;
+        const void * kernel;
+
+        bool operator==(const cache_key & other) const { return device == other.device && kernel == other.kernel; }
+    };
+
+    struct cache_key_hash {
+        // MurmurHash3 mixing function for better hash distribution (vs. just std::hash which in some implementations simply returns the identity)
+        static size_t hash_mix(size_t x) {
+            std::uint64_t       y = x;
+            const std::uint64_t m = 0xe9846af9b1a615d;
+
+            y ^= y >> 32;
+            y *= m;
+            y ^= y >> 32;
+            y *= m;
+            y ^= y >> 28;
+
+            return static_cast<size_t>(y);
+        }
+
+        size_t operator()(const cache_key & key) const {
+            // Use a nonzero seed to avoid mapping all-zero keys to zero
+            size_t h = 42;
+            h        = hash_mix(h + key.device);
+            h        = hash_mix(h + reinterpret_cast<size_t>(key.kernel));
+            return h;
+        }
+    };
+
+    static std::mutex                                          cache_mutex;
+    static std::unordered_map<cache_key, bool, cache_key_hash> cache;
+
+    const cache_key             key = { device, kernel };
+    std::lock_guard<std::mutex> lock(cache_mutex);
+    const auto                  it = cache.find(key);
+    if (it != cache.end()) {
+        return it->second;
+    }
+
+    cudaFuncAttributes attr = {};
+    CUDA_CHECK(cudaFuncGetAttributes(&attr, kernel));
+
+    // PDL device-side primitives are emitted only for PTX versions >= 90.
+    // We have to guard on a loaded kernel's PTX version so a kernel forward-JIT'ed
+    // from pre-Hopper PTX to a Hopper-or-newer GPU does not opt into PDL.
+    const bool can_use_pdl = attr.ptxVersion >= 90;
+    cache.emplace(key, can_use_pdl);
+    return can_use_pdl;
+}
+
 #endif //defined(GGML_CUDA_USE_PDL)
 
 
@@ -1561,8 +1621,7 @@ static __inline__ void ggml_cuda_kernel_launch(Kernel kernel, const ggml_cuda_ke
         return env == nullptr || std::atoi(env) != 0;
     }();
 
-    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
-    if (env_pdl_enabled && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_HOPPER) {
+    if (env_pdl_enabled && ggml_cuda_kernel_can_use_pdl(reinterpret_cast<const void *>(kernel))) {
         auto pdl_cfg = ggml_cuda_pdl_config(launch_params);
 
         CUDA_CHECK(cudaLaunchKernelEx(&pdl_cfg.cfg, kernel, std::forward<Args>(args)... ));

ggml/src/ggml-cuda/fattn-common.cuh

@@ -1153,8 +1153,8 @@ void launch_fattn(
 
     GGML_ASSERT(block_dim.x % warp_size == 0);
 
-    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(blocks_num, block_dim, nbytes_shared, main_stream);
-    ggml_cuda_kernel_launch(fattn_kernel, launch_params,
+        // disabled PDL enrollment for now due to a compiler bug.
+        fattn_kernel<<<blocks_num, block_dim, nbytes_shared, main_stream>>>(
         (const char *) Q->data,
         K_data,
         V_data,

ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -472,7 +472,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_mask(
 
             const int i = 8 * (threadIdx.x % (nbatch_fa/8));
 
-            cp_async_cg_16<preload>(tile_mask_32 + j_sram*(nbatch_fa*sizeof(half) + 16) + i*sizeof(half), mask_h + j_vram*stride_mask + i);
+            cp_async_cg_16<preload>(tile_mask_32 + j_sram*(nbatch_fa*sizeof(half) + 16) + i*sizeof(half), mask_h + int64_t(j_vram)*stride_mask + i);
         }
     } else if constexpr (oob_check) {
 #pragma unroll
@@ -488,7 +488,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_mask(
             for (int i0 = 0; i0 < nbatch_fa; i0 += warp_size) {
                 const int i = i0 + threadIdx.x;
 
-                tile_mask[j_sram*(nbatch_fa + 8) + i] = i < i_sup ? mask_h[j_vram*stride_mask + i] : half(0.0f);
+                tile_mask[j_sram*(nbatch_fa + 8) + i] = i < i_sup ? mask_h[int64_t(j_vram)*stride_mask + i] : half(0.0f);
             }
         }
     } else if constexpr (nbatch_fa < 2*warp_size) {
@@ -505,7 +505,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_mask(
 
             const int i = threadIdx.x % (warp_size/cols_per_warp);
 
-            ggml_cuda_memcpy_1<sizeof(half2)>(tile_mask + j_sram*(nbatch_fa + 8) + 2*i, mask_h + j_vram*stride_mask + 2*i);
+            ggml_cuda_memcpy_1<sizeof(half2)>(tile_mask + j_sram*(nbatch_fa + 8) + 2*i, mask_h + int64_t(j_vram)*stride_mask + 2*i);
         }
     } else {
 #pragma unroll
@@ -521,7 +521,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_mask(
             for (int i0 = 0; i0 < nbatch_fa; i0 += 2*warp_size) {
                 const int i = i0 + 2*threadIdx.x;
 
-                ggml_cuda_memcpy_1<sizeof(half2)>(tile_mask + j_sram*(nbatch_fa + 8) + i, mask_h + j_vram*stride_mask + i);
+                ggml_cuda_memcpy_1<sizeof(half2)>(tile_mask + j_sram*(nbatch_fa + 8) + i, mask_h + int64_t(j_vram)*stride_mask + i);
             }
         }
     }

ggml/src/ggml-cuda/fwht.cu

@@ -0,0 +1,101 @@
+#include "common.cuh"
+#include "fwht.cuh"
+
+template <int N>
+__launch_bounds__(4*ggml_cuda_get_physical_warp_size(), 1)
+__global__ void fwht_cuda(const float * src, float * dst, const int64_t n_rows, const float scale) {
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+
+    const int64_t r = (int64_t) blockIdx.x * blockDim.y + threadIdx.y;
+
+    if (r >= n_rows) {
+        return;
+    }
+
+    src += r * N;
+    dst += r * N;
+
+    static constexpr int el_w = N / warp_size;
+    float     reg[el_w];
+    const int lane = threadIdx.x;
+
+    ggml_cuda_pdl_sync();
+#pragma unroll
+    for (int i = 0; i < el_w; ++i) {
+        reg[i] = src[i * warp_size + lane] * scale;
+    }
+
+#pragma unroll
+    for (int h = 1; h < warp_size; h *= 2) {
+#pragma unroll
+        for (int j = 0; j < el_w; j++) {
+            const float val  = reg[j];
+            const float val2 = __shfl_xor_sync(0xFFFFFFFF, val, h, warp_size);
+
+            reg[j] = (lane & h) == 0 ? val + val2 : val2 - val;
+        }
+    }
+
+#pragma unroll
+    for (int h = warp_size; h < N; h *= 2) {
+        const int step = h / warp_size;
+#pragma unroll
+        for (int j = 0; j < el_w; j += 2 * step) {
+#pragma unroll
+            for (int k = 0; k < step; k++) {
+                const float x = reg[j + k];
+                const float y = reg[j + k + step];
+
+                reg[j + k]        = x + y;
+                reg[j + k + step] = x - y;
+            }
+        }
+    }
+
+#pragma unroll
+    for (int i = 0; i < el_w; ++i) {
+        dst[i * warp_size + lane] = reg[i];
+    }
+}
+
+bool ggml_cuda_op_fwht(ggml_backend_cuda_context & ctx, const ggml_tensor * src, ggml_tensor * dst) {
+    GGML_ASSERT(ggml_are_same_shape(src, dst));
+    if (!ggml_is_contiguous(src) || !ggml_is_contiguous(dst)) {
+        return false;
+    }
+    const int     n    = src->ne[0];
+    const int64_t rows = ggml_nrows(src);
+
+    const float * src_d = (const float *) src->data;
+    float *       dst_d = (float *) dst->data;
+
+    const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size;
+    const int rows_per_block = 4;
+
+    const int64_t num_blocks = (rows + rows_per_block - 1) / rows_per_block;
+
+    cudaStream_t                         stream = ctx.stream();
+    dim3                                 grid_dims(num_blocks, 1, 1);
+    dim3                                 block_dims(warp_size, rows_per_block, 1);
+    const ggml_cuda_kernel_launch_params launch_params =
+        ggml_cuda_kernel_launch_params(grid_dims, block_dims, 0, stream);
+
+    const float scale = 1 / sqrtf(n);
+
+    switch (n) {
+        case 64:
+            ggml_cuda_kernel_launch(fwht_cuda<64>, launch_params, src_d, dst_d, rows, scale);
+            return true;
+        case 128:
+            ggml_cuda_kernel_launch(fwht_cuda<128>, launch_params, src_d, dst_d, rows, scale);
+            return true;
+        case 256:
+            ggml_cuda_kernel_launch(fwht_cuda<256>, launch_params, src_d, dst_d, rows, scale);
+            return true;
+        case 512:
+            ggml_cuda_kernel_launch(fwht_cuda<512>, launch_params, src_d, dst_d, rows, scale);
+            return true;
+        default:
+            return false;
+    }
+}

ggml/src/ggml-cuda/fwht.cuh

@@ -0,0 +1,4 @@
+#include "common.cuh"
+
+// Returns whether the Fast Walsh-Hadamard transform could be used.
+bool ggml_cuda_op_fwht(ggml_backend_cuda_context & ctx, const ggml_tensor * src, ggml_tensor * dst);

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -24,6 +24,7 @@
 #include "ggml-cuda/diagmask.cuh"
 #include "ggml-cuda/diag.cuh"
 #include "ggml-cuda/fattn.cuh"
+#include "ggml-cuda/fwht.cuh"
 #include "ggml-cuda/getrows.cuh"
 #include "ggml-cuda/im2col.cuh"
 #include "ggml-cuda/mmf.cuh"
@@ -2569,6 +2570,7 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
             use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1], /*n_experts=*/0);
             use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src0->nb, src1->ne[1], /*mul_mat_id=*/false);
             use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src0->nb, src1->ne[1]);
+            use_mul_mat_vec_q       = use_mul_mat_vec_q         && ggml_cuda_should_use_mmvq(src0->type, cc, src1->ne[1]);
             any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
         }
     } else {
@@ -2577,6 +2579,7 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
         use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1], /*n_experts=*/0);
         use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src0->nb, src1->ne[1], /*mul_mat_id=*/false);
         use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src0->nb, src1->ne[1]);
+        use_mul_mat_vec_q       = use_mul_mat_vec_q         && ggml_cuda_should_use_mmvq(src0->type, cc, src1->ne[1]);
         any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
     }
 
@@ -2594,6 +2597,11 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     bool use_batched_cublas_bf16 = src0->type == GGML_TYPE_BF16 && bf16_mma_hardware_available(cc);
     bool use_batched_cublas_f32  = src0->type == GGML_TYPE_F32;
 
+    const int32_t hint = ggml_get_op_params_i32(dst, 1);
+    if (hint == GGML_HINT_SRC0_IS_HADAMARD && !split && ggml_cuda_op_fwht(ctx, src1, dst)) {
+        return;
+    }
+
     if (!split && use_mul_mat_vec_f) {
         // the custom F16 vector kernel can be used over batched cuBLAS GEMM
         // but this is only faster for GPUs without tensor cores or with a thin src0 matrix (particularly KQV in attention)
@@ -4986,8 +4994,14 @@ static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t *
 }
 
 static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend_dev_t dev) {
-    GGML_UNUSED(dev);
-    return GGML_BACKEND_DEVICE_TYPE_GPU;
+    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *) dev->context;
+
+    cudaDeviceProp prop;
+    CUDA_CHECK(cudaGetDeviceProperties(&prop, ctx->device));
+
+    return prop.integrated
+        ? GGML_BACKEND_DEVICE_TYPE_IGPU
+        : GGML_BACKEND_DEVICE_TYPE_GPU;
 }
 
 static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {

ggml/src/ggml-cuda/mmvq.cu

@@ -63,6 +63,7 @@ static constexpr __host__ __device__ int get_vdr_mmvq(ggml_type type) {
 
 enum mmvq_parameter_table_id {
     MMVQ_PARAMETERS_GENERIC = 0,
+    MMVQ_PARAMETERS_TURING,
     MMVQ_PARAMETERS_GCN,
     MMVQ_PARAMETERS_RDNA2,
     MMVQ_PARAMETERS_RDNA3_0,
@@ -78,6 +79,8 @@ static constexpr __device__ mmvq_parameter_table_id get_device_table_id() {
     return MMVQ_PARAMETERS_RDNA2;
 #elif defined(GCN) || defined(CDNA)
     return MMVQ_PARAMETERS_GCN;
+#elif defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING && __CUDA_ARCH__ < GGML_CUDA_CC_AMPERE
+    return MMVQ_PARAMETERS_TURING;
 #else
     return MMVQ_PARAMETERS_GENERIC;
 #endif
@@ -96,6 +99,9 @@ static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
     if (GGML_CUDA_CC_IS_GCN(cc) || GGML_CUDA_CC_IS_CDNA(cc)) {
         return MMVQ_PARAMETERS_GCN;
     }
+    if (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_TURING && ggml_cuda_highest_compiled_arch(cc) < GGML_CUDA_CC_AMPERE) {
+        return MMVQ_PARAMETERS_TURING;
+    }
     return MMVQ_PARAMETERS_GENERIC;
 }
 
@@ -271,6 +277,53 @@ int get_mmvq_mmid_max_batch(ggml_type type, int cc) {
     return MMVQ_MAX_BATCH_SIZE;
 }
 
+bool ggml_cuda_should_use_mmvq(enum ggml_type type, int cc, int64_t ne11) {
+    if (GGML_CUDA_CC_IS_CDNA(cc)) {
+        if (GGML_CUDA_CC_IS_CDNA1(cc)) {
+            switch (type) {
+                case GGML_TYPE_Q4_0:
+                case GGML_TYPE_Q4_1:
+                    return ne11 <= 7;
+                case GGML_TYPE_Q5_1:
+                    return ne11 <= 7;
+                case GGML_TYPE_Q8_0:
+                    return ne11 <= 6;
+                case GGML_TYPE_Q2_K:
+                    return ne11 <= 4;
+                case GGML_TYPE_Q3_K:
+                    return ne11 <= 3;
+                case GGML_TYPE_Q4_K:
+                    return ne11 <= 2;
+                case GGML_TYPE_Q5_K:
+                    return ne11 <= 3;
+                case GGML_TYPE_Q6_K:
+                    return ne11 <= 4;
+                case GGML_TYPE_IQ1_S:
+                    return ne11 <= 5;
+                case GGML_TYPE_IQ2_XXS:
+                case GGML_TYPE_IQ3_S:
+                case GGML_TYPE_IQ4_XS:
+                    return ne11 <= 6;
+                default:
+                    return ne11 <= MMVQ_MAX_BATCH_SIZE;
+            }
+        }
+        switch (type) { // tuned for CDNA2
+            case GGML_TYPE_Q2_K:
+                return ne11 <= 5;
+            case GGML_TYPE_Q3_K:
+            case GGML_TYPE_Q4_K:
+            case GGML_TYPE_Q5_K:
+                return ne11 <= 3;
+            case GGML_TYPE_Q6_K:
+                return ne11 <= 5;
+            default:
+                return ne11 <= MMVQ_MAX_BATCH_SIZE;
+        }
+    }
+    return ne11 <= MMVQ_MAX_BATCH_SIZE;
+}
+
 // Device constexpr: returns the max batch size for the current arch+type at compile time.
 template <ggml_type type>
 static constexpr __device__ int get_mmvq_mmid_max_batch_for_device() {
@@ -370,11 +423,38 @@ static constexpr __host__ __device__ int calc_nwarps(ggml_type type, int ncols_d
         }
         return 1;
     }
+    if (table_id == MMVQ_PARAMETERS_TURING) {
+        if (ncols_dst == 1) {
+            switch (type) {
+                case GGML_TYPE_Q2_K:
+                case GGML_TYPE_Q3_K:
+                case GGML_TYPE_Q4_K:
+                case GGML_TYPE_Q5_K:
+                case GGML_TYPE_Q6_K:
+                    return 2;
+                default:
+                    return 4;
+            }
+        }
+        switch (ncols_dst) {
+            case 2:
+            case 3:
+            case 4:
+                return 4;
+            case 5:
+            case 6:
+            case 7:
+            case 8:
+                return 2;
+            default:
+                return 1;
+        }
+    }
     return 1;
 }
 
 static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int table_id, bool small_k = false, int nwarps = 1) {
-    if (table_id == MMVQ_PARAMETERS_GENERIC || table_id == MMVQ_PARAMETERS_GCN) {
+    if (table_id == MMVQ_PARAMETERS_GENERIC || table_id == MMVQ_PARAMETERS_GCN || table_id == MMVQ_PARAMETERS_TURING) {
         switch (ncols_dst) {
             case 1:
                 return small_k ? nwarps : 1;

ggml/src/ggml-cuda/mmvq.cuh

@@ -2,6 +2,8 @@
 
 #define MMVQ_MAX_BATCH_SIZE 8 // Max. batch size for which to use MMVQ kernels.
 
+bool ggml_cuda_should_use_mmvq(enum ggml_type type, int cc, int64_t ne11);
+
 // Returns the maximum batch size for which MMVQ should be used for MUL_MAT_ID,
 // based on the quantization type and GPU architecture (compute capability).
 int get_mmvq_mmid_max_batch(ggml_type type, int cc);

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -39,7 +39,7 @@
 #include "ggml-hexagon.h"
 #include "ggml-impl.h"
 #include "ggml-quants.h"
-#include "op-desc.h"
+#include "htp-opnode.h"
 #include "htp-ops.h"
 #include "htp_iface.h"
 #include "htp-drv.h"
@@ -68,6 +68,7 @@ static u32vec opt_pmu_evt { 0x3, 0x111, 0x100, 0x105, 0x240, 0x256, 0x7D, 0x8C }
 static int opt_opstage  = HTP_OPSTAGE_QUEUE | HTP_OPSTAGE_COMPUTE;
 static int opt_opbatch  = 1024; // max number of ops in a batch
 static int opt_opqueue  = 16;   // max number of pending batches
+static int opt_oppoll   = 0;    // polling for batch completions
 
 static std::regex* opt_opfilter = NULL; // regex of ops to not claim
 
@@ -101,23 +102,23 @@ static const char * status_to_str(uint32_t status) {
 
 // ** debug helpers
 
-static void ggml_hexagon_dump_op_exec(const std::string &sess_name, const ggml_tensor * op, const uint32_t req_flags) {
+static void ggml_hexagon_dump_op_exec(const std::string &sess_name, const htp_opnode & node, const uint32_t req_flags) {
     if (!opt_verbose) return;
 
-    op_desc desc(op);
+    htp_opformat fmt(node);
     GGML_LOG_DEBUG("ggml-hex: %s execute-op %s: %s : %s : %s : %s : %s : flags 0x%x\n", sess_name.c_str(),
-                ggml_op_desc(op), desc.names, desc.dims, desc.types, desc.strides, desc.buffs, req_flags);
+                node.op_name().c_str(), fmt.names, fmt.dims, fmt.types, fmt.strides, fmt.buffs, req_flags);
 }
 
 static void ggml_hexagon_dump_op_supp(const std::string &sess_name, const struct ggml_tensor * op, bool supp) {
     if (!opt_verbose) return;
 
-    op_desc desc(op);
+    htp_opformat fmt(htp_opformat(htp_opnode{const_cast<ggml_tensor*>(op), {}, HTP_OP_INVALID}));
     GGML_LOG_DEBUG("ggml-hex: %s supports-op %s: %s : %s : %s : %s : %s : %s\n", sess_name.c_str(),
-                ggml_op_desc(op), desc.names, desc.dims, desc.types, desc.strides, desc.buffs, supp ? "yes" : "no");
+                ggml_op_desc(op), fmt.names, fmt.dims, fmt.types, fmt.strides, fmt.buffs, supp ? "yes" : "no");
 }
 
-static void ggml_hexagon_dump_op_prof(const std::string &sess_name, const ggml_tensor * op,
+static void ggml_hexagon_dump_op_prof(const std::string &sess_name, const htp_opnode & node,
                                       uint32_t op_usec, uint32_t op_cycles, const uint32_t pmu[]) {
     if (!opt_profile) return;
 
@@ -128,15 +129,16 @@ static void ggml_hexagon_dump_op_prof(const std::string &sess_name, const ggml_t
                 pmu[0], pmu[1], pmu[2], pmu[3], pmu[4], pmu[5], pmu[6], pmu[7]);
     }
 
-    op_desc desc(op);
+    htp_opformat fmt(node);
     GGML_LOG_DEBUG("ggml-hex: %s profile-op %s: %s : %s : %s : %s : usec %u cycles %u%s\n", sess_name.c_str(),
-            ggml_op_desc(op), desc.names, desc.dims, desc.types, desc.strides, op_usec, op_cycles, pmu_str);
+            node.op_name().c_str(), fmt.names, fmt.dims, fmt.types, fmt.strides, op_usec, op_cycles, pmu_str);
 }
 
 // ** backend sessions
 
 struct ggml_hexagon_opbatch;
 struct ggml_hexagon_opqueue;
+struct htp_opnode;
 
 struct ggml_hexagon_session {
     std::string      name;
@@ -166,7 +168,7 @@ struct ggml_hexagon_session {
     void allocate(int dev_id) noexcept(false);
     void release() noexcept(true);
 
-    void enqueue_op(htp_op_code opcode, const ggml_tensor *op);
+    void enqueue_op(const htp_opnode & node);
     void flush(bool all = true);
 
     void flush_pending(bool all = false);
@@ -550,7 +552,7 @@ static void repack_q4_0_q4x4x2(ggml_tensor * t, const void * data, size_t size)
 
     size_t row_size    = ggml_row_size(t->type, t->ne[0]);
     size_t row_size_pd = ggml_row_size(t->type, hex_round_up(t->ne[0], QK_Q4_0x4x2));  // extra elements for the pad
-    size_t row_size_rp = row_size * 2;  // extra space for tmp pad (if any)
+    size_t row_size_rp = row_size_pd;  // scratch must hold one full padded tile (qblk_size/2 quants + scales)
 
     // Ensure we don't try to read more data than is available in the source buffer 'data'
     // or write more than the tensor can hold.
@@ -611,7 +613,7 @@ static void repack_q4x4x2_q4_0(void * data, const ggml_tensor * t, size_t size)
 
     size_t row_size    = ggml_row_size(t->type, t->ne[0]);
     size_t row_size_pd = ggml_row_size(t->type, hex_round_up(t->ne[0], QK_Q4_0x4x2));  // extra elements for the pad
-    size_t row_size_rp = row_size * 2;  // extra space for tmp pad (if any)
+    size_t row_size_rp = row_size_pd;  // scratch must hold one full padded tile (qblk_size/2 quants + scales)
 
     // Ensure we don't try to copy more data than the tensor actually contains.
     const size_t total_tensor_size = (size_t)nrows * row_size;
@@ -660,6 +662,239 @@ static void repack_q4x4x2_q4_0(void * data, const ggml_tensor * t, size_t size)
     ggml_aligned_free(buf_rp, row_size_rp);
 }
 
+static void unpack_q4_1_quants(uint8_t * qs, const block_q4_1 * x, unsigned int bi) {
+    static const int qk = QK4_1;
+
+    for (unsigned int i = 0; i < qk / 2; ++i) {
+        const int x0             = (x->qs[i] & 0x0F);
+        const int x1             = (x->qs[i] >> 4);
+        qs[bi * qk + i + 0]      = x0;
+        qs[bi * qk + i + qk / 2] = x1;
+    }
+}
+
+static void pack_q4_1_quants(block_q4_1 * x, const uint8_t * qs, unsigned int bi) {
+    static const int qk = QK4_1;
+
+    for (unsigned int i = 0; i < qk / 2; ++i) {
+        const uint8_t x0 = qs[bi * qk + i + 0];
+        const uint8_t x1 = qs[bi * qk + i + qk / 2];
+        x->qs[i]         = x0 | (x1 << 4);
+    }
+}
+
+static void repack_row_q4_1x4x2(uint8_t * y, const block_q4_1 * x, int64_t k) {
+    static const int qk = QK_Q4_0x4x2;
+    const int        nb = (k + qk - 1) / qk;  // number of blocks (padded)
+    const int        nloe = k % qk;           // leftovers
+
+    const int dblk_size = 8 * 4;              // 8x (d, m) __fp16 = 32 bytes
+    const int qblk_size = qk / 2;             // int4 = 128 bytes
+    const int qrow_size = k / 2;              // int4 (not padded to blocks)
+
+    uint8_t * y_q = y + 0;                    // quants first
+    uint8_t * y_d = y + qrow_size;            // then scales/offsets
+
+    // Repack the quants
+    for (int i = 0; i < nb; i++) {
+        uint8_t qs[QK_Q4_0x4x2];  // unpacked quants
+        unpack_q4_1_quants(qs, &x[i * 8 + 0], 0);
+        unpack_q4_1_quants(qs, &x[i * 8 + 1], 1);
+        unpack_q4_1_quants(qs, &x[i * 8 + 2], 2);
+        unpack_q4_1_quants(qs, &x[i * 8 + 3], 3);
+        unpack_q4_1_quants(qs, &x[i * 8 + 4], 4);
+        unpack_q4_1_quants(qs, &x[i * 8 + 5], 5);
+        unpack_q4_1_quants(qs, &x[i * 8 + 6], 6);
+        unpack_q4_1_quants(qs, &x[i * 8 + 7], 7);
+
+        bool partial = (nloe && i == nb-1);
+
+        uint8_t * q = y_q + (i * qblk_size);
+        for (int j = 0; j < qk / 2; j++) {
+            q[j] = partial ? (qs[j*2+1] << 4) | qs[j*2+0] : (qs[j+128] << 4) | qs[j+000];
+        }
+    }
+
+    // Repack the scales and offsets
+    for (int i = 0; i < nb; i++) {
+        ggml_half * d_m = (ggml_half *) (y_d + i * dblk_size);
+        for (int j = 0; j < 8; j++) {
+            d_m[j * 2 + 0] = x[i * 8 + j].d;
+            d_m[j * 2 + 1] = x[i * 8 + j].m;
+        }
+    }
+}
+
+static void unpack_row_q4_1x4x2(block_q4_1 * x, const uint8_t * y, int64_t k) {
+    static const int qk = QK_Q4_0x4x2;
+    const int        nb = (k + qk - 1) / qk;  // number of blocks (padded)
+    const int        nloe = k % qk;           // leftovers
+
+    const int dblk_size = 8 * 4;              // 8x (d, m) __fp16 = 32 bytes
+    const int qblk_size = qk / 2;             // int4 = 128 bytes
+    const int qrow_size = k / 2;              // int4 (not padded to blocks)
+
+    const uint8_t * y_q = y + 0;              // quants first
+    const uint8_t * y_d = y + qrow_size;      // then scales/offsets
+
+    // Unpack the quants
+    for (int i = 0; i < nb; i++) {
+        uint8_t qs[QK_Q4_0x4x2];
+        bool partial = (nloe && i == nb-1);
+
+        const uint8_t * q = y_q + (i * qblk_size);
+        for (int j = 0; j < qk / 2; j++) {
+            if (partial) {
+                qs[j*2+0] = q[j] & 0x0F;
+                qs[j*2+1] = q[j] >> 4;
+            } else {
+                qs[j+000] = q[j] & 0x0F;
+                qs[j+128] = q[j] >> 4;
+            }
+        }
+
+        pack_q4_1_quants(&x[i * 8 + 0], qs, 0);
+        pack_q4_1_quants(&x[i * 8 + 1], qs, 1);
+        pack_q4_1_quants(&x[i * 8 + 2], qs, 2);
+        pack_q4_1_quants(&x[i * 8 + 3], qs, 3);
+        pack_q4_1_quants(&x[i * 8 + 4], qs, 4);
+        pack_q4_1_quants(&x[i * 8 + 5], qs, 5);
+        pack_q4_1_quants(&x[i * 8 + 6], qs, 6);
+        pack_q4_1_quants(&x[i * 8 + 7], qs, 7);
+    }
+
+    // Unpack the scales and offsets
+    for (int i = 0; i < nb; i++) {
+        const ggml_half * d_m = (const ggml_half *) (y_d + i * dblk_size);
+        for (int j = 0; j < 8; j++) {
+            x[i * 8 + j].d = d_m[j * 2 + 0];
+            x[i * 8 + j].m = d_m[j * 2 + 1];
+        }
+    }
+}
+
+static void init_row_q4_1x4x2(block_q4_1 * x, int64_t k) {
+    static const int qk = QK_Q4_0x4x2;
+    const int        nb = (k + qk - 1) / qk;  // number of blocks (padded)
+
+    uint8_t qs[QK_Q4_0x4x2];  // unpacked quants
+    memset(qs, 0, sizeof(qs));
+
+    for (int i = 0; i < nb; i++) {
+        pack_q4_1_quants(&x[i * 8 + 0], qs, 0);
+        pack_q4_1_quants(&x[i * 8 + 1], qs, 1);
+        pack_q4_1_quants(&x[i * 8 + 2], qs, 2);
+        pack_q4_1_quants(&x[i * 8 + 3], qs, 3);
+        pack_q4_1_quants(&x[i * 8 + 4], qs, 4);
+        pack_q4_1_quants(&x[i * 8 + 5], qs, 5);
+        pack_q4_1_quants(&x[i * 8 + 6], qs, 6);
+        pack_q4_1_quants(&x[i * 8 + 7], qs, 7);
+    }
+
+    for (int i = 0; i < nb; i++) {
+        for (int j = 0; j < 8; j++) {
+            x[i * 8 + j].d = 0;
+            x[i * 8 + j].m = 0;
+        }
+    }
+}
+
+static void repack_q4_1_q4x4x2(ggml_tensor * t, const void * data, size_t size) {
+    int64_t nrows = ggml_nrows(t);
+
+    size_t row_size    = ggml_row_size(t->type, t->ne[0]);
+    size_t row_size_pd = ggml_row_size(t->type, hex_round_up(t->ne[0], QK_Q4_0x4x2));
+    size_t row_size_rp = row_size_pd;  // scratch must hold one full padded tile (qblk_size/2 quants + scales)
+
+    const size_t total_tensor_size = (size_t)nrows * row_size;
+    const size_t n_bytes_to_copy = size < total_tensor_size ? size : total_tensor_size;
+
+    const int64_t n_full_rows = n_bytes_to_copy / row_size;
+    const size_t  n_rem_bytes = n_bytes_to_copy % row_size;
+
+    void * buf_pd = ggml_aligned_malloc(row_size_pd);
+    GGML_ASSERT(buf_pd != NULL);
+
+    void * buf_rp = ggml_aligned_malloc(row_size_rp);
+    GGML_ASSERT(buf_rp != NULL);
+
+    HEX_VERBOSE("ggml-hex: repack-q4_1-q4x4x2 %s : data %p size %zu dims %ldx%ld row-size %zu\n", t->name, data, size,
+                t->ne[0], nrows, row_size);
+
+    init_row_q4_1x4x2((block_q4_1 *) buf_pd, t->ne[0]);
+
+    for (int64_t i = 0; i < n_full_rows; i++) {
+        const uint8_t * src = (const uint8_t *) data + (i * row_size);
+        uint8_t *       dst = (uint8_t *) t->data + (i * row_size);
+
+        memcpy(buf_pd, src, row_size);
+        repack_row_q4_1x4x2((uint8_t *) buf_rp, (const block_q4_1 *) buf_pd, t->ne[0]);
+        memcpy(dst, buf_rp, row_size);
+    }
+
+    if (n_rem_bytes > 0) {
+        const int64_t i = n_full_rows;
+        const uint8_t * src = (const uint8_t *) data + (i * row_size);
+        uint8_t *       dst = (uint8_t *) t->data + (i * row_size);
+
+        init_row_q4_1x4x2((block_q4_1 *) buf_pd, t->ne[0]);
+        memcpy(buf_pd, src, n_rem_bytes);
+        repack_row_q4_1x4x2((uint8_t *) buf_rp, (const block_q4_1 *) buf_pd, t->ne[0]);
+        memcpy(dst, buf_rp, n_rem_bytes);
+    }
+
+    ggml_aligned_free(buf_pd, row_size_pd);
+    ggml_aligned_free(buf_rp, row_size_rp);
+}
+
+static void repack_q4x4x2_q4_1(void * data, const ggml_tensor * t, size_t size) {
+    int64_t nrows = ggml_nrows(t);
+
+    size_t row_size    = ggml_row_size(t->type, t->ne[0]);
+    size_t row_size_pd = ggml_row_size(t->type, hex_round_up(t->ne[0], QK_Q4_0x4x2));
+    size_t row_size_rp = row_size_pd;  // scratch must hold one full padded tile (qblk_size/2 quants + scales)
+
+    const size_t total_tensor_size = (size_t)nrows * row_size;
+    const size_t n_bytes_to_copy = size < total_tensor_size ? size : total_tensor_size;
+
+    const int64_t n_full_rows = n_bytes_to_copy / row_size;
+    const size_t  n_rem_bytes = n_bytes_to_copy % row_size;
+
+    void * buf_pd = ggml_aligned_malloc(row_size_pd);
+    GGML_ASSERT(buf_pd != NULL);
+
+    void * buf_rp = ggml_aligned_malloc(row_size_rp);
+    GGML_ASSERT(buf_rp != NULL);
+
+    HEX_VERBOSE("ggml-hex: repack-q4x4x2-q4_1 %s : data %p size %zu dims %ldx%ld row-size %zu\n", t->name, data, size,
+                t->ne[0], nrows, row_size);
+
+    memset(buf_rp, 0, row_size_rp);  // clear-out padded buffer to make sure the tail is all zeros
+
+    for (int64_t i = 0; i < n_full_rows; i++) {
+        const uint8_t * src = (const uint8_t *) t->data + (i * row_size);
+        uint8_t *       dst = (uint8_t *) data + (i * row_size);
+
+        memcpy(buf_rp, src, row_size);
+        unpack_row_q4_1x4x2((block_q4_1 *) buf_pd, (const uint8_t *) buf_rp, t->ne[0]);
+        memcpy(dst, buf_pd, row_size);
+    }
+
+    if (n_rem_bytes > 0) {
+        const int64_t i = n_full_rows;
+        const uint8_t * src = (const uint8_t *) t->data + (i * row_size);
+        uint8_t *       dst = (uint8_t *) data + (i * row_size);
+
+        // We still need to read and unpack the entire source row because quantization is block-based.
+        memcpy(buf_rp, src, row_size);
+        unpack_row_q4_1x4x2((block_q4_1 *) buf_pd, (const uint8_t *) buf_rp, t->ne[0]);
+        memcpy(dst, buf_pd, n_rem_bytes);
+    }
+
+    ggml_aligned_free(buf_pd, row_size_pd);
+    ggml_aligned_free(buf_rp, row_size_rp);
+}
+
 // ======== Q8x4x2 ====================
 static void dump_block_q8_0(const block_q8_0 * b, int i) {
     HEX_VERBOSE("ggml-hex: repack q8_0 %d: %d %d %d %d ... %d %d %d %d : %.6f\n", i, b->qs[0], b->qs[1], b->qs[2],
@@ -876,7 +1111,7 @@ static void repack_q8_0_q8x4x2(ggml_tensor * t, const void * data, size_t size)
 
     size_t row_size    = ggml_row_size(t->type, t->ne[0]);
     size_t row_size_pd = ggml_row_size(t->type, hex_round_up(t->ne[0], QK_Q8_0x4x2));  // extra elements for the pad
-    size_t row_size_rp = row_size * 2;  // extra space for tmp pad (if any)
+    size_t row_size_rp = row_size_pd;  // scratch must hold one full padded tile (qblk_size quants + scales)
 
     // Ensure we don't try to read more data than is available in the source buffer 'data'
     // or write more than the tensor can hold.
@@ -937,7 +1172,7 @@ static void repack_q8x4x2_q8_0(void * data, const ggml_tensor * t, size_t size)
 
     size_t row_size    = ggml_row_size(t->type, t->ne[0]);
     size_t row_size_pd = ggml_row_size(t->type, hex_round_up(t->ne[0], QK_Q8_0x4x2));  // extra elements for the pad
-    size_t row_size_rp = row_size * 2;  // extra space for tmp pad (if any)
+    size_t row_size_rp = row_size_pd;  // scratch must hold one full padded tile (qblk_size quants + scales)
 
     // Ensure we don't try to copy more data than the tensor actually contains.
     const size_t total_tensor_size = (size_t)nrows * row_size;
@@ -1238,7 +1473,7 @@ static void repack_mxfp4_mxfp4x4x2(ggml_tensor * t, const void * data, size_t si
 
     size_t row_size    = ggml_row_size(t->type, t->ne[0]);
     size_t row_size_pd = ggml_row_size(t->type, hex_round_up(t->ne[0], QK_MXFP4x4x2));  // extra elements for the pad
-    size_t row_size_rp = row_size * 2;  // extra space for tmp pad (if any)
+    size_t row_size_rp = row_size_pd;  // scratch must hold one full padded tile (qblk_size/2 quants + scales)
 
     // Ensure we don't try to read more data than is available in the source buffer 'data'
     // or write more than the tensor can hold.
@@ -1299,7 +1534,7 @@ static void repack_mxfp4x4x2_mxfp4(void * data, const ggml_tensor * t, size_t si
 
     size_t row_size    = ggml_row_size(t->type, t->ne[0]);
     size_t row_size_pd = ggml_row_size(t->type, hex_round_up(t->ne[0], QK_MXFP4x4x2));  // extra elements for the pad
-    size_t row_size_rp = row_size * 2;  // extra space for tmp pad (if any)
+    size_t row_size_rp = row_size_pd;  // scratch must hold one full padded tile (qblk_size/2 quants + scales)
 
     // Ensure we don't try to copy more data than the tensor actually contains.
     const size_t total_tensor_size = (size_t)nrows * row_size;
@@ -1365,6 +1600,12 @@ static void ggml_backend_hexagon_buffer_set_tensor(ggml_backend_buffer_t buffer,
             repack_q4_0_q4x4x2(tensor, data, size);
             break;
 
+        case GGML_TYPE_Q4_1:
+            GGML_ASSERT(offset == 0);
+            GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
+            repack_q4_1_q4x4x2(tensor, data, size);
+            break;
+
         case GGML_TYPE_Q8_0:
             GGML_ASSERT(offset == 0);
             GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
@@ -1407,6 +1648,12 @@ static void ggml_backend_hexagon_buffer_get_tensor(ggml_backend_buffer_t buffer,
             repack_q4x4x2_q4_0(data, tensor, size);
             break;
 
+        case GGML_TYPE_Q4_1:
+            GGML_ASSERT(offset == 0);
+            GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
+            repack_q4x4x2_q4_1(data, tensor, size);
+            break;
+
         case GGML_TYPE_Q8_0:
             GGML_ASSERT(offset == 0);
             GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
@@ -1536,12 +1783,10 @@ static ggml_backend_buffer_type_i ggml_backend_hexagon_repack_buffer_type_interf
     /* .is_host          = */ ggml_backend_hexagon_repack_buffer_type_is_host,
 };
 
-// Backend session implementation
-
 struct ggml_hexagon_opbatch {
     ggml_hexagon_session*            sess;
 
-    std::vector<const ggml_tensor*>  ops;       // pointers to original ops
+    std::vector<htp_opnode>          ops;       // htp_opnode of ops
 
     std::vector<htp_buf_desc>        h_bufs;    // htp buffer descriptors
     std::vector<htp_tensor>          h_tens;    // htp tensor descriptors
@@ -1673,7 +1918,7 @@ struct ggml_hexagon_opbatch {
         return ti;
     }
 
-    bool fit_op(const struct ggml_tensor *t) const {
+    bool fit_op(const htp_opnode & node) const {
         if (n_ops >= n_ops_max ) return false;
 
         // check how much extras we will need
@@ -1693,10 +1938,10 @@ struct ggml_hexagon_opbatch {
             }
         };
 
-        for (unsigned int i=0; i < HTP_OP_MAX_INPUTS && t->src[i]; i++) {
-            fit_tensor(t->src[i]);
+        for (const auto * src : node.get_inputs()) {
+            fit_tensor(src);
         }
-        fit_tensor(t);
+        fit_tensor(node.dst());
 
         if ((extra_bufs + n_bufs) > n_bufs_max) return false;
         if ((extra_tens + n_tens) > n_tens_max) return false;
@@ -1706,29 +1951,30 @@ struct ggml_hexagon_opbatch {
     }
 
     // assumes that fit_op() was called first and returned true
-    void add_op(htp_op_code opcode, const struct ggml_tensor * t) {
+    void add_op(const htp_opnode & node) {
         // Add new op
 
         unsigned int n = n_ops++;
         GGML_ASSERT(n_ops <= n_ops_max);
 
-        ops[n] = t;
+        ops[n] = node;
 
         htp_op_desc &o = h_ops[n];
-        memcpy(&o.params, &t->op_params, sizeof(t->op_params));
-        o.opcode = opcode;
+        memcpy(&o.params, &node.node->op_params, sizeof(node.node->op_params));
+        o.opcode = node.opcode;
         o.flags  = 0;
 
         if (!(opt_opstage & HTP_OPSTAGE_COMPUTE)) {
             o.flags |= HTP_OPFLAGS_SKIP_COMPUTE;
         }
 
-        ggml_hexagon_dump_op_exec(sess->c_name(), t, o.flags);
+        ggml_hexagon_dump_op_exec(sess->c_name(), node, o.flags);
 
+        auto inputs = node.get_inputs();
         for (unsigned int i=0; i < HTP_OP_MAX_INPUTS; i++) {
-            o.src[i] = t->src[i] ? add_tensor(t->src[i]) : 0xffff;
+            o.src[i] = (i < inputs.size() && inputs[i]) ? add_tensor(inputs[i]) : 0xffff;
         }
-        o.dst = add_tensor(t);
+        o.dst = add_tensor(node.dst());
     }
 };
 
@@ -1737,7 +1983,7 @@ struct ggml_hexagon_opqueue {
     ggml_hexagon_shared_buffer *shm_buf;
     size_t                      shm_blk_size;
 
-    using opvec = std::vector<const ggml_tensor*>;
+    using opvec = std::vector<htp_opnode>;
 
     std::queue<unsigned int>    done;       // completed batch ids
     std::vector<opvec>          op_cache;   // per batch op cache
@@ -1886,7 +2132,8 @@ void ggml_hexagon_session::flush_pending(bool all) {
         uint32_t               n_dbufs;
 
         // Read response packet from queue
-        int err = dspqueue_read(this->queue, &flags, 1, &n_dbufs, &dbuf, sizeof(rsp), &rsp_size, (uint8_t *) &rsp, DSPQUEUE_TIMEOUT);
+        const uint32_t timeo = opt_oppoll ? 0 : DSPQUEUE_TIMEOUT;
+        int err = dspqueue_read(this->queue, &flags, 1, &n_dbufs, &dbuf, sizeof(rsp), &rsp_size, (uint8_t *) &rsp, timeo);
         if (err == AEE_EEXPIRED) {
             continue;
         }
@@ -1935,11 +2182,11 @@ void ggml_hexagon_session::flush_batch() {
     }
 }
 
-void ggml_hexagon_session::enqueue_op(htp_op_code opcode, const ggml_tensor *op) {
-    if (!op_batch->fit_op(op)) {
+void ggml_hexagon_session::enqueue_op(const htp_opnode & node) {
+    if (!op_batch->fit_op(node)) {
         flush_batch();
     }
-    op_batch->add_op(opcode, op);
+    op_batch->add_op(node);
 }
 
 // Flush HTP response queue i.e wait for all outstanding requests to complete
@@ -2290,6 +2537,7 @@ static bool ggml_hexagon_supported_gated_delta_net(const struct ggml_hexagon_ses
     const int64_t H        = v->ne[1];
     const int64_t n_tokens = v->ne[2];
     const int64_t n_seqs   = v->ne[3];
+    const int64_t K        = state->ne[1];
 
     if (S_v <= 0 || S_v > 128 || H <= 0 || n_tokens <= 0 || n_seqs <= 0) {
         return false;
@@ -2302,10 +2550,10 @@ static bool ggml_hexagon_supported_gated_delta_net(const struct ggml_hexagon_ses
     if ((g->ne[0] != 1 && g->ne[0] != S_v) || beta->ne[0] != 1) {
         return false;
     }
-    if (ggml_nelements(state) != S_v * S_v * H * n_seqs) {
+    if (ggml_nelements(state) != S_v * S_v * H * n_seqs * K) {
         return false;
     }
-    if (dst->ne[0] != S_v * H || dst->ne[1] != n_tokens * n_seqs + S_v * n_seqs) {
+    if (dst->ne[0] != S_v * H || dst->ne[1] != n_tokens * n_seqs + S_v * n_seqs * K) {
         return false;
     }
 
@@ -2327,6 +2575,7 @@ static bool ggml_hexagon_supported_mul_mat(const struct ggml_hexagon_session * s
 
     switch (src0->type) {
         case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_MXFP4:
@@ -2377,6 +2626,7 @@ static bool ggml_hexagon_supported_mul_mat_id(const struct ggml_hexagon_session
 
     switch (src0->type) {
         case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_MXFP4:
@@ -2874,6 +3124,7 @@ static htp_op_code op_remap_to_htp(const ggml_tensor * t) {
         case GGML_OP_NORM:            return HTP_OP_NORM;
         case GGML_OP_L2_NORM:         return HTP_OP_L2_NORM;
         case GGML_OP_RMS_NORM:        return HTP_OP_RMS_NORM;
+        case GGML_OP_CONCAT:          return HTP_OP_CONCAT;
         case GGML_OP_SCALE:           return HTP_OP_SCALE;
         case GGML_OP_SQR:             return HTP_OP_SQR;
         case GGML_OP_SQRT:            return HTP_OP_SQRT;
@@ -2928,10 +3179,43 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
 
     HEX_VERBOSE("ggml-hex: %s graph-compute n_nodes %d\n", sess->c_name(), graph->n_nodes);
 
+    std::vector<htp_opnode> nodes;
+    nodes.reserve(graph->n_nodes);
+
+    // Fusion
     for (int i = 0; i < graph->n_nodes; ++i) {
         ggml_tensor * n = graph->nodes[i];
-        if (op_is_compute(n) && (opt_opstage & HTP_OPSTAGE_QUEUE)) {
-            sess->enqueue_op(op_remap_to_htp(n), n);
+        if (!op_is_compute(n)) {
+            continue;
+        }
+
+        ggml_tensor * next_node = (i + 1 < graph->n_nodes) ? graph->nodes[i + 1] : nullptr;
+
+        htp_opnode node = {
+            /*.node =*/ n,
+            /*.fused =*/ {},
+            /*.opcode =*/ HTP_OP_INVALID
+        };
+
+        if (n->op == GGML_OP_RMS_NORM && next_node) {
+            if (next_node->op == GGML_OP_MUL && op_is_compute(next_node) && ggml_can_fuse(graph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
+                node.add_fused(next_node);
+                node.opcode = HTP_OP_RMS_NORM_MUL;
+                i++; // skip the fused MUL node
+            }
+        }
+
+        if (node.opcode == HTP_OP_INVALID) {
+            node.opcode = op_remap_to_htp(n);
+        }
+
+        nodes.push_back(std::move(node));
+    }
+
+    // Queue and execute
+    if (opt_opstage & HTP_OPSTAGE_QUEUE) {
+        for (const auto & node : nodes) {
+            sess->enqueue_op(node);
         }
     }
 
@@ -2950,51 +3234,7 @@ static void ggml_backend_hexagon_synchronize(ggml_backend_t backend) {
     sess->flush();
 }
 
-struct node_info {
-    ggml_tensor * node;
-
-    std::vector<ggml_tensor *> fused;
-
-    ggml_op op() const {
-        return node->op;
-    }
-
-    const ggml_tensor * dst() const {
-        return fused.empty() ? node : fused.back();
-    }
-
-    const ggml_tensor * src0() const {
-        return node->src[0];
-    }
-
-    const ggml_tensor * src1() const {
-        return node->src[1];
-    }
-
-    bool is_empty() const {
-        return ggml_op_is_empty(node->op);
-    }
-
-    void add_fused(ggml_tensor * t) {
-        fused.push_back(t);
-    }
-
-    bool stackable() const {
-        switch (this->op()) {
-            case GGML_OP_MUL_MAT:
-            case GGML_OP_MUL_MAT_ID:
-                return ggml_is_quantized(this->src0()->type);
-            default:
-                return false;
-        }
-    }
-
-    bool same_input(const node_info& n) const {
-        return n.src1() == this->src1();
-    }
-};
-
-static std::vector<int> ggml_hexagon_graph_optimize_reorder(const std::vector<node_info> & nodes) {
+static std::vector<int> ggml_hexagon_graph_optimize_reorder(const std::vector<htp_opnode> & nodes) {
     const int n = nodes.size();
 
     std::vector<int> res;
@@ -3048,14 +3288,14 @@ static void ggml_backend_hexagon_graph_optimize(ggml_backend_t backend, ggml_cgr
 
     enum ggml_op ops[MAX_FUSE];
 
-    std::vector<node_info> nodes;
+    std::vector<htp_opnode> nodes;
     nodes.reserve(gf->n_nodes);
 
     // fuse nodes:
     // we don't want to make reorders that break fusing, so we first pack all fusable tensors
     //   and perform the reorder over the fused nodes. after the reorder is done, we unfuse
     for (int i = 0; i < n; i++) {
-        node_info node = {
+        htp_opnode node = {
             /*.node =*/gf->nodes[i],
             /*.fused =*/{},
         };
@@ -3286,6 +3526,25 @@ static bool ggml_hexagon_supported_repeat(const struct ggml_hexagon_session * se
     return true;
 }
 
+static bool ggml_hexagon_supported_concat(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
+    int dim = ((const int32_t *) op->op_params)[0];
+    if (dim < 0 || dim >= GGML_MAX_DIMS) {
+        return false;
+    }
+
+    for (int i = 0; i < GGML_MAX_SRC; ++i) {
+        const struct ggml_tensor * src = op->src[i];
+        if (!src) {
+            continue;
+        }
+        if (src->type != GGML_TYPE_F32 && src->type != GGML_TYPE_I32 && src->type != GGML_TYPE_F16) {
+            return false;
+        }
+    }
+
+    return true;
+}
+
 static bool ggml_hexagon_supported_fill(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
     const struct ggml_tensor * dst = op;
 
@@ -3434,6 +3693,10 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
             supp = ggml_hexagon_supported_cumsum(sess, op);
             break;
 
+        case GGML_OP_CONCAT:
+            supp = ggml_hexagon_supported_concat(sess, op);
+            break;
+
         case GGML_OP_FILL:
             supp = ggml_hexagon_supported_fill(sess, op);
             break;
@@ -3598,6 +3861,8 @@ static void ggml_hexagon_init(ggml_backend_reg * reg) {
     // Basic sanity checks to make sure definitions match
     static_assert((unsigned int) HTP_TYPE_Q4_0 == (unsigned int) GGML_TYPE_Q4_0,
                   "please update hexagon_type to match ggml_type");
+    static_assert((unsigned int) HTP_TYPE_Q4_1 == (unsigned int) GGML_TYPE_Q4_1,
+                  "please update hexagon_type to match ggml_type");
     static_assert((unsigned int) HTP_TYPE_Q8_0 == (unsigned int) GGML_TYPE_Q8_0,
                   "please update hexagon_type to match ggml_type");
     static_assert((unsigned int) HTP_TYPE_MXFP4 == (unsigned int) GGML_TYPE_MXFP4,
@@ -3610,6 +3875,7 @@ static void ggml_hexagon_init(ggml_backend_reg * reg) {
     const char * str_opstage  = getenv("GGML_HEXAGON_OPSTAGE");
     const char * str_opbatch  = getenv("GGML_HEXAGON_OPBATCH");
     const char * str_opqueue  = getenv("GGML_HEXAGON_OPQUEUE");
+    const char * str_oppoll   = getenv("GGML_HEXAGON_OPPOLL");
     const char * str_opfilter = getenv("GGML_HEXAGON_OPFILTER");
     const char * str_profile  = getenv("GGML_HEXAGON_PROFILE");
     const char * str_etm      = getenv("GGML_HEXAGON_ETM");
@@ -3647,6 +3913,7 @@ static void ggml_hexagon_init(ggml_backend_reg * reg) {
     opt_opstage   = str_opstage  ? strtoul(str_opstage, NULL, 0)          : opt_opstage;
     opt_opbatch   = str_opbatch  ? strtoul(str_opbatch, NULL, 0)          : opt_opbatch;
     opt_opqueue   = str_opqueue  ? strtoul(str_opqueue, NULL, 0)          : opt_opqueue;
+    opt_oppoll    = str_oppoll   ? strtoul(str_oppoll,  NULL, 0)          : opt_oppoll;
     opt_profile   = str_profile  ? atoi(str_profile)                      : 0;
     opt_etm       = str_etm      ? atoi(str_etm)                          : 0;
     opt_nhvx      = str_nhvx     ? strtoul(str_nhvx, NULL, 0)             : opt_nhvx;

ggml/src/ggml-hexagon/htp-opnode.h

@@ -0,0 +1,241 @@
+#ifndef HTP_OPNODE_H
+#define HTP_OPNODE_H
+
+#define GGML_COMMON_IMPL_CPP
+#include "ggml-backend-impl.h"
+#include "ggml-common.h"
+
+#include <string>
+#include <vector>
+#include <stdio.h>
+#include "htp-ops.h"
+
+struct htp_opnode {
+    ggml_tensor * node = nullptr;
+
+    std::vector<ggml_tensor *> fused;
+
+    htp_op_code opcode = HTP_OP_INVALID;
+
+    ggml_op op() const {
+        return node->op;
+    }
+
+    const ggml_tensor * dst() const {
+        return fused.empty() ? node : fused.back();
+    }
+
+    const ggml_tensor * src0() const {
+        return node->src[0];
+    }
+
+    const ggml_tensor * src1() const {
+        return node->src[1];
+    }
+
+    bool is_empty() const {
+        return ggml_op_is_empty(node->op);
+    }
+
+    void add_fused(ggml_tensor * t) {
+        fused.push_back(t);
+    }
+
+    bool stackable() const {
+        switch (this->op()) {
+            case GGML_OP_MUL_MAT:
+            case GGML_OP_MUL_MAT_ID:
+                return ggml_is_quantized(this->src0()->type);
+            default:
+                return false;
+        }
+    }
+
+    bool same_input(const htp_opnode& n) const {
+        return n.src1() == this->src1();
+    }
+
+    std::vector<const ggml_tensor *> get_inputs() const {
+        std::vector<const ggml_tensor *> inputs;
+        std::vector<const ggml_tensor *> outputs;
+        outputs.push_back(node);
+        for (const auto * f : fused) {
+            outputs.push_back(f);
+        }
+
+        auto contains = [&](const std::vector<const ggml_tensor *> & vec, const ggml_tensor * t) {
+            for (const auto * x : vec) {
+                if (x == t) return true;
+            }
+            return false;
+        };
+
+        auto add_input = [&](const ggml_tensor * t) {
+            if (t && !contains(outputs, t) && !contains(inputs, t)) {
+                inputs.push_back(t);
+            }
+        };
+
+        for (int i = 0; i < GGML_MAX_SRC && node->src[i]; i++) {
+            add_input(node->src[i]);
+        }
+        for (const auto * f : fused) {
+            for (int i = 0; i < GGML_MAX_SRC && f->src[i]; i++) {
+                add_input(f->src[i]);
+            }
+        }
+        return inputs;
+    }
+
+    std::string op_name() const {
+        if (fused.empty()) {
+            return ggml_op_desc(node);
+        }
+        std::string name = ggml_op_desc(node);
+        for (const auto * f : fused) {
+            name += "+";
+            name += ggml_op_desc(f);
+        }
+        return name;
+    }
+};
+
+struct htp_opformat {
+    char strides[64 * GGML_MAX_SRC];
+    char dims[64 * GGML_MAX_SRC];
+    char types[16 * GGML_MAX_SRC];
+    char buffs[64 * GGML_MAX_SRC];
+    char names[64 * GGML_MAX_SRC];
+
+    int format_tensor_dims(char * str, const struct ggml_tensor * t) {
+        if (t->ne[2] == 1 && t->ne[3] == 1) {
+            return sprintf(str, "%d:%d", (int) t->ne[0], (int) t->ne[1]);
+        } else {
+            return sprintf(str, "%d:%d:%d:%d", (int) t->ne[0], (int) t->ne[1], (int) t->ne[2], (int) t->ne[3]);
+        }
+    }
+
+    void format_op_dims(char * str, const htp_opnode & node) {
+        char * p = str;
+        auto inputs = node.get_inputs();
+
+        if (!inputs.empty()) {
+            p += format_tensor_dims(p, inputs[0]);
+
+            for (size_t i = 1; i < inputs.size(); i++) {
+                p += sprintf(p, " x ");
+                p += format_tensor_dims(p, inputs[i]);
+            }
+
+            p += sprintf(p, " -> ");
+        }
+
+        char self[64];
+        format_tensor_dims(self, node.dst());
+        p += sprintf(p, "%s", self);
+    }
+
+    int format_tensor_strides(char * str, const struct ggml_tensor * t) {
+        const char * c = ggml_is_contiguous(t) ? "" : "!";
+
+        if (t->ne[2] == 1 && t->ne[3] == 1) {
+            return sprintf(str, "%zu:%zu%s", (size_t) t->nb[0], (size_t) t->nb[1], c);
+        } else {
+            return sprintf(str, "%zu:%zu:%zu:%zu%s", (size_t) t->nb[0], (size_t) t->nb[1], (size_t) t->nb[2], (size_t) t->nb[3], c);
+        }
+    }
+
+    void format_op_strides(char * str, const htp_opnode & node) {
+        char * p = str;
+        auto inputs = node.get_inputs();
+
+        if (!inputs.empty()) {
+            p += format_tensor_strides(p, inputs[0]);
+
+            for (size_t i = 1; i < inputs.size(); i++) {
+                p += sprintf(p, " x ");
+                p += format_tensor_strides(p, inputs[i]);
+            }
+
+            p += sprintf(p, " -> ");
+        }
+
+        char self[64];
+        format_tensor_strides(self, node.dst());
+        p += sprintf(p, "%s", self);
+    }
+
+    void format_op_types(char * str, const htp_opnode & node) {
+        char * p = str;
+        auto inputs = node.get_inputs();
+
+        if (!inputs.empty()) {
+            p += sprintf(p, "%s", ggml_type_name(inputs[0]->type));
+
+            for (size_t i = 1; i < inputs.size(); i++) {
+                p += sprintf(p, " x ");
+                p += sprintf(p, "%s", ggml_type_name(inputs[i]->type));
+            }
+
+            p += sprintf(p, " -> ");
+        }
+
+        p += sprintf(p, "%s", ggml_type_name(node.dst()->type));
+    }
+
+    const char * tensor_buff_name(const struct ggml_tensor * t) {
+        if (t->buffer) {
+            return ggml_backend_buffer_name(t->buffer);
+        }
+        return "NONE";
+    }
+
+    void format_op_buffs(char * str, const htp_opnode & node) {
+        char * p = str;
+        auto inputs = node.get_inputs();
+
+        if (!inputs.empty()) {
+            p += sprintf(p, "%s", tensor_buff_name(inputs[0]));
+
+            for (size_t i = 1; i < inputs.size(); i++) {
+                p += sprintf(p, " x ");
+                p += sprintf(p, "%s", tensor_buff_name(inputs[i]));
+            }
+
+            p += sprintf(p, " -> ");
+        }
+
+        p += sprintf(p, "%s", tensor_buff_name(node.dst()));
+    }
+
+    void format_op_names(char * str, const htp_opnode & node) {
+        char * p = str;
+        auto inputs = node.get_inputs();
+
+        if (!inputs.empty()) {
+            p += sprintf(p, "%s", inputs[0]->name);
+
+            for (size_t i = 1; i < inputs.size(); i++) {
+                p += sprintf(p, " x ");
+                p += sprintf(p, "%s", inputs[i]->name);
+            }
+
+            p += sprintf(p, " -> ");
+        }
+
+        p += sprintf(p, "%s", node.dst()->name);
+    }
+
+    void format(const htp_opnode & node) {
+        format_op_dims(dims, node);
+        format_op_strides(strides, node);
+        format_op_types(types, node);
+        format_op_buffs(buffs, node);
+        format_op_names(names, node);
+    }
+
+    htp_opformat() {}
+    htp_opformat(const htp_opnode & node) { format(node); }
+};
+
+#endif // HTP_OPNODE_H

ggml/src/ggml-hexagon/htp/CMakeLists.txt

@@ -35,6 +35,7 @@ add_library(${HTP_LIB} SHARED
     ssm-conv.c
     cumsum-ops.c
     fill-ops.c
+    concat-ops.c
     diag-ops.c
     solve-tri-ops.c
     gated-delta-net-ops.c
@@ -57,15 +58,16 @@ list(FIND HTP_HMX_VERSIONS ${DSP_VERSION} _hmx_idx)
 
 if (_hmx_idx GREATER_EQUAL 0)
     target_sources(${HTP_LIB} PRIVATE
-        hmx-queue.c
-        hmx-matmul-ops.c
         hmx-flash-attn-ops.c
+        hmx-matmul-ops.c
+        hmx-queue.c
     )
 
     # -mhmx enables HMX instruction set (needed by files that include hmx-utils.h)
     set_source_files_properties(
-        hmx-matmul-ops.c
         hmx-flash-attn-ops.c
+        hmx-matmul-ops.c
+        hmx-queue.c
         PROPERTIES COMPILE_OPTIONS "-mhmx"
     )
 

ggml/src/ggml-hexagon/htp/concat-ops.c

@@ -0,0 +1,275 @@
+#include "htp-ctx.h"
+#include "htp-ops.h"
+#include "hexagon_types.h"
+#include "hexagon_protos.h"
+#include "hvx_hexagon_protos.h"
+#include "hex-dma.h"
+#include "vtcm-utils.h"
+#include "hvx-utils.h"
+#include "hex-fastdiv.h"
+#include <string.h>
+
+struct htp_concat_context {
+    struct htp_ops_context * octx;
+    uint32_t dim;
+    uint32_t nrows_per_thread;
+    struct fastdiv_values div_ne0;
+    struct fastdiv_values div_ne1;
+    struct fastdiv_values div_ne2;
+};
+
+static void concat_2d_f32_transposed(unsigned int nth, unsigned int ith, void * data) {
+    struct htp_concat_context * cctx = (struct htp_concat_context *) data;
+    struct htp_ops_context * octx = cctx->octx;
+
+    const struct htp_tensor * src0 = octx->src[0];
+    const struct htp_tensor * src1 = octx->src[1];
+    const struct htp_tensor * dst  = octx->dst;
+
+    const uint32_t src0_ne0 = src0->ne[0];
+    const uint32_t src1_ne0 = src1->ne[0];
+    const uint32_t ne1      = dst->ne[1];
+
+    const uint32_t start_i = ith * cctx->nrows_per_thread;
+    const uint32_t end_i   = (start_i + cctx->nrows_per_thread < ne1) ? (start_i + cctx->nrows_per_thread) : ne1;
+    if (start_i >= end_i) return;
+
+    dma_queue * q = octx->ctx->dma[ith];
+
+    uint8_t * spad0_base = octx->src0_spad.data + ith * octx->src0_spad.size_per_thread;
+    uint8_t * spad1_base = octx->src1_spad.data + ith * octx->src1_spad.size_per_thread;
+
+    const uint32_t block_i = 32;
+    const uint32_t spad1_stride = block_i * sizeof(float);
+
+    int32_t offsets[32] __attribute__((aligned(128)));
+    for(int k=0; k<32; k++) {
+        offsets[k] = k * spad1_stride;
+    }
+    HVX_Vector vv = *(HVX_Vector*)offsets;
+    const uint32_t src1_ne0_padded = hex_round_up(src1_ne0, 32);
+    const uint32_t spad0_row_bytes = hex_round_up((src0_ne0 + src1_ne0_padded) * sizeof(float), VLEN);
+    uint32_t mu = src1_ne0_padded * spad1_stride;
+
+    for (uint32_t i = start_i; i < end_i; i += block_i) {
+        uint32_t current_block_i = (end_i - i < block_i) ? (end_i - i) : block_i;
+
+        uint32_t src1_width_bytes = current_block_i * sizeof(float);
+        uint8_t * src1_ptr = (uint8_t *)src1->data + i * src1->nb[1];
+        dma_queue_push(q, dma_make_ptr(spad1_base, src1_ptr), spad1_stride, src1->nb[0], src1_width_bytes, src1_ne0);
+
+        uint32_t src0_row_bytes = src0_ne0 * sizeof(float);
+        uint8_t * src0_ptr = (uint8_t *)src0->data + i * src0->nb[1];
+        dma_queue_push(q, dma_make_ptr(spad0_base, src0_ptr), spad0_row_bytes, src0->nb[1], src0_row_bytes, current_block_i);
+
+        dma_queue_pop(q); // src1
+
+        HVX_Vector * vtcm_tmp = (HVX_Vector *)(spad1_base + src1_ne0_padded * spad1_stride);
+
+        for (uint32_t j = 0; j < src1_ne0_padded; j += 32) {
+            #pragma unroll(4)
+            for (uint32_t ii = 0; ii < current_block_i; ii++) {
+                size_t rt = (size_t)(spad1_base + j * spad1_stride + ii * sizeof(float));
+                Q6_vgather_ARMVw(&vtcm_tmp[ii], rt, mu, vv);
+                uint8_t * dst_ptr = spad0_base + ii * spad0_row_bytes + (src0_ne0 + j) * sizeof(float);
+                hvx_vmemu(dst_ptr) = vtcm_tmp[ii];
+            }
+        }
+
+        dma_queue_pop(q); // src0
+
+        uint8_t * dst_ptr = (uint8_t *)dst->data + i * dst->nb[1];
+        dma_queue_push(q, dma_make_ptr(dst_ptr, spad0_base), dst->nb[1], spad0_row_bytes, (src0_ne0 + src1_ne0) * sizeof(float), current_block_i);
+
+        dma_queue_pop(q);
+    }
+}
+
+static void concat_2d_f16_transposed(unsigned int nth, unsigned int ith, void * data) {
+    struct htp_concat_context * cctx = (struct htp_concat_context *) data;
+    struct htp_ops_context * octx = cctx->octx;
+
+    const struct htp_tensor * src0 = octx->src[0];
+    const struct htp_tensor * src1 = octx->src[1];
+    const struct htp_tensor * dst  = octx->dst;
+
+    const uint32_t src0_ne0 = src0->ne[0];
+    const uint32_t src1_ne0 = src1->ne[0];
+    const uint32_t ne1      = dst->ne[1];
+
+    const uint32_t start_i = ith * cctx->nrows_per_thread;
+    const uint32_t end_i   = (start_i + cctx->nrows_per_thread < ne1) ? (start_i + cctx->nrows_per_thread) : ne1;
+    if (start_i >= end_i) return;
+
+    dma_queue * q = octx->ctx->dma[ith];
+
+    uint8_t * spad0_base = octx->src0_spad.data + ith * octx->src0_spad.size_per_thread;
+    uint8_t * spad1_base = octx->src1_spad.data + ith * octx->src1_spad.size_per_thread;
+
+    const uint32_t block_i = 64;
+    const uint32_t spad1_stride = block_i * sizeof(__fp16);
+
+    int16_t offsets[64] __attribute__((aligned(128)));
+    for(int k=0; k<64; k++) {
+        offsets[k] = k * spad1_stride;
+    }
+    HVX_Vector vv = *(HVX_Vector*)offsets;
+    const uint32_t src1_ne0_padded = hex_round_up(src1_ne0, 64);
+    const uint32_t spad0_row_bytes = hex_round_up((src0_ne0 + src1_ne0_padded) * sizeof(__fp16), VLEN);
+    uint32_t mu = src1_ne0_padded * spad1_stride;
+
+    for (uint32_t i = start_i; i < end_i; i += block_i) {
+        uint32_t current_block_i = (end_i - i < block_i) ? (end_i - i) : block_i;
+
+        uint32_t src1_width_bytes = current_block_i * sizeof(__fp16);
+        uint8_t * src1_ptr = (uint8_t *)src1->data + i * src1->nb[1];
+        dma_queue_push(q, dma_make_ptr(spad1_base, src1_ptr), spad1_stride, src1->nb[0], src1_width_bytes, src1_ne0);
+
+        uint32_t src0_row_bytes = src0_ne0 * sizeof(__fp16);
+        uint8_t * src0_ptr = (uint8_t *)src0->data + i * src0->nb[1];
+        dma_queue_push(q, dma_make_ptr(spad0_base, src0_ptr), spad0_row_bytes, src0->nb[1], src0_row_bytes, current_block_i);
+
+        dma_queue_pop(q); // src1
+
+        HVX_Vector * vtcm_tmp = (HVX_Vector *)(spad1_base + src1_ne0_padded * spad1_stride);
+
+        for (uint32_t j = 0; j < src1_ne0_padded; j += 64) {
+            #pragma unroll(4)
+            for (uint32_t ii = 0; ii < current_block_i; ii++) {
+                size_t rt = (size_t)(spad1_base + j * spad1_stride + ii * sizeof(__fp16));
+                Q6_vgather_ARMVh(&vtcm_tmp[ii], rt, mu, vv);
+                uint8_t * dst_ptr = spad0_base + ii * spad0_row_bytes + (src0_ne0 + j) * sizeof(__fp16);
+                hvx_vmemu(dst_ptr) = vtcm_tmp[ii];
+            }
+        }
+
+        dma_queue_pop(q); // src0
+
+        uint8_t * dst_ptr = (uint8_t *)dst->data + i * dst->nb[1];
+        dma_queue_push(q, dma_make_ptr(dst_ptr, spad0_base), dst->nb[1], spad0_row_bytes, (src0_ne0 + src1_ne0) * sizeof(__fp16), current_block_i);
+
+        dma_queue_pop(q);
+    }
+}
+
+static void concat_generic(unsigned int nth, unsigned int ith, void * data) {
+    struct htp_concat_context * cctx = (struct htp_concat_context *) data;
+    struct htp_ops_context * octx = cctx->octx;
+
+    const struct htp_tensor * src0 = octx->src[0];
+    const struct htp_tensor * src1 = octx->src[1];
+    const struct htp_tensor * dst  = octx->dst;
+
+    const int dim = cctx->dim;
+    const uint32_t type_size = (dst->type == HTP_TYPE_F32 || dst->type == HTP_TYPE_I32) ? 4 : 2;
+
+    const uint32_t ne[4] = {dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3]};
+    const uint32_t total_elements = ne[0] * ne[1] * ne[2] * ne[3];
+    const uint32_t chunk_size = (total_elements + nth - 1) / nth;
+
+    const uint32_t start_idx = MIN(ith * chunk_size, total_elements);
+    const uint32_t end_idx   = MIN(start_idx + chunk_size, total_elements);
+
+    // Naive scalar element-wise copy
+    for (uint32_t idx = start_idx; idx < end_idx; idx++) {
+        uint32_t idx_div_ne0 = fastdiv(idx, &cctx->div_ne0);
+        uint32_t i0 = idx - idx_div_ne0 * ne[0];
+
+        uint32_t idx_div_ne01 = fastdiv(idx_div_ne0, &cctx->div_ne1);
+        uint32_t i1 = idx_div_ne0 - idx_div_ne01 * ne[1];
+
+        uint32_t idx_div_ne012 = fastdiv(idx_div_ne01, &cctx->div_ne2);
+        uint32_t i2 = idx_div_ne01 - idx_div_ne012 * ne[2];
+        uint32_t i3 = idx_div_ne012;
+
+        uint8_t * dst_ptr = (uint8_t *)dst->data + i3 * dst->nb[3] + i2 * dst->nb[2] + i1 * dst->nb[1] + i0 * dst->nb[0];
+
+        uint32_t idx_dim = 0;
+        if (dim == 0) idx_dim = i0;
+        else if (dim == 1) idx_dim = i1;
+        else if (dim == 2) idx_dim = i2;
+        else if (dim == 3) idx_dim = i3;
+
+        const struct htp_tensor * src = (idx_dim < src0->ne[dim]) ? src0 : src1;
+
+        uint32_t s0 = i0;
+        uint32_t s1 = i1;
+        uint32_t s2 = i2;
+        uint32_t s3 = i3;
+
+        if (dim == 0 && src == src1) s0 -= src0->ne[0];
+        if (dim == 1 && src == src1) s1 -= src0->ne[1];
+        if (dim == 2 && src == src1) s2 -= src0->ne[2];
+        if (dim == 3 && src == src1) s3 -= src0->ne[3];
+
+        uint8_t * src_ptr = (uint8_t *)src->data + s3 * src->nb[3] + s2 * src->nb[2] + s1 * src->nb[1] + s0 * src->nb[0];
+
+        if (type_size == 4) {
+            *(float*)dst_ptr = *(float*)src_ptr;
+        } else {
+            *(__fp16*)dst_ptr = *(__fp16*)src_ptr;
+        }
+    }
+}
+
+int op_concat(struct htp_ops_context * octx) {
+    const struct htp_tensor * src0 = octx->src[0];
+    const struct htp_tensor * src1 = octx->src[1];
+    const struct htp_tensor * dst  = octx->dst;
+
+    int dim = octx->op_params[0];
+
+    bool is_2d = dst->ne[2] == 1 && dst->ne[3] == 1;
+
+    const uint32_t type_size = (dst->type == HTP_TYPE_F32 || dst->type == HTP_TYPE_I32) ? 4 : 2;
+    bool is_src1_transposed  = (src1->nb[0] > src1->nb[1]);
+    bool is_src0_transposed  = (src0->nb[0] > src0->nb[1]);
+
+    uint32_t n_threads = octx->n_threads;
+    struct htp_concat_context cctx;
+    cctx.octx = octx;
+    cctx.dim = dim;
+    cctx.div_ne0 = init_fastdiv_values(dst->ne[0]);
+    cctx.div_ne1 = init_fastdiv_values(dst->ne[1]);
+    cctx.div_ne2 = init_fastdiv_values(dst->ne[2]);
+
+    void (*worker_func)(unsigned int, unsigned int, void *) = concat_generic;
+
+    if (dim == 0 && is_2d && is_src1_transposed && !is_src0_transposed) {
+        n_threads = MIN(dst->ne[1], n_threads);
+        if (n_threads < 1) {
+            n_threads = 1;
+        }
+        uint32_t block_i = (type_size == 4) ? 32 : 64;
+
+        cctx.nrows_per_thread = hmx_ceil_div(dst->ne[1], n_threads);
+
+        // Allocate VTCM
+        uint32_t spad1_stride = block_i * type_size;
+
+        uint32_t src1_ne0_padded = hex_round_up(src1->ne[0], block_i);
+        uint32_t spad0_row_bytes = hex_round_up((src0->ne[0] + src1_ne0_padded) * type_size, VLEN);
+
+        octx->src0_spad.size_per_thread = block_i * spad0_row_bytes;
+        octx->src1_spad.size_per_thread = src1_ne0_padded * spad1_stride + block_i * VLEN;
+
+        octx->src0_spad.size = n_threads * octx->src0_spad.size_per_thread;
+        octx->src1_spad.size = n_threads * octx->src1_spad.size_per_thread;
+
+        if (octx->src0_spad.size + octx->src1_spad.size > octx->ctx->vtcm_size) {
+            return HTP_STATUS_VTCM_TOO_SMALL;
+        }
+
+        octx->src0_spad.data = octx->ctx->vtcm_base;
+        octx->src1_spad.data = octx->src0_spad.data + octx->src0_spad.size;
+
+        if (type_size == 4) {
+            worker_func = concat_2d_f32_transposed;
+        } else {
+            worker_func = concat_2d_f16_transposed;
+        }
+    }
+
+    worker_pool_run_func(octx->ctx->worker_pool, worker_func, &cctx, n_threads);
+    return HTP_STATUS_OK;
+}

ggml/src/ggml-hexagon/htp/cpy-ops.c

@@ -28,158 +28,170 @@ struct htp_copy_context {
     uint32_t          dst_blocks_per_row;
 
     uint32_t          src0_nrows_per_thread;
-
-    void (*copy)(struct htp_copy_context * ct, struct htp_ops_context * octx, int nth, int ith);
 };
 
 #define cpy_preamble                              \
     const struct htp_tensor *src0 = octx->src[0]; \
     const struct htp_tensor *dst  = octx->dst;    \
                                                   \
-    const uint32_t ne00 = src0->ne[0];     \
-    const uint32_t ne01 = src0->ne[1];     \
-    const uint32_t ne02 = src0->ne[2];     \
-    const uint32_t ne03 = src0->ne[3];     \
-                                           \
-    const uint32_t nb00 = src0->nb[0];     \
-    const uint32_t nb01 = src0->nb[1];     \
-    const uint32_t nb02 = src0->nb[2];     \
-    const uint32_t nb03 = src0->nb[3];     \
-                                           \
-    const uint32_t  ne0 = dst->ne[0];      \
-    const uint32_t  ne1 = dst->ne[1];      \
-    const uint32_t  ne2 = dst->ne[2];      \
-    const uint32_t  ne3 = dst->ne[3];      \
-                                           \
-    const uint32_t  nb0 = dst->nb[0];      \
-    const uint32_t  nb1 = dst->nb[1];      \
-    const uint32_t  nb2 = dst->nb[2];      \
-    const uint32_t  nb3 = dst->nb[3];      \
-                                           \
+    const uint32_t ne00 = src0->ne[0];            \
+    const uint32_t ne01 = src0->ne[1];            \
+    const uint32_t ne02 = src0->ne[2];            \
+    const uint32_t ne03 = src0->ne[3];            \
+                                                  \
+    const uint32_t nb00 = src0->nb[0];            \
+    const uint32_t nb01 = src0->nb[1];            \
+    const uint32_t nb02 = src0->nb[2];            \
+    const uint32_t nb03 = src0->nb[3];            \
+                                                  \
+    const uint32_t  ne0 = dst->ne[0];             \
+    const uint32_t  ne1 = dst->ne[1];             \
+    const uint32_t  ne2 = dst->ne[2];             \
+    const uint32_t  ne3 = dst->ne[3];             \
+                                                  \
+    const uint32_t  nb0 = dst->nb[0];             \
+    const uint32_t  nb1 = dst->nb[1];             \
+    const uint32_t  nb2 = dst->nb[2];             \
+    const uint32_t  nb3 = dst->nb[3];             \
+                                                  \
     const uint32_t   nr = ne01;
 
-static void cpy_thread_sametype_sameshape(struct htp_copy_context * ct, struct htp_ops_context * octx, const int nth, const int ith) {
-    cpy_preamble;
-
-    // parallelize by src0 rows
-    const uint32_t dr  = ct->src0_nrows_per_thread;
-    const uint32_t ir0 = dr * ith;
-    const uint32_t ir1 = (ir0 + dr) < nr ? (ir0 + dr) : nr;
-
-    // copy by rows
-    for (uint32_t i03 = 0; i03 < ne03; i03++) {
-        for (uint32_t i02 = 0; i02 < ne02; i02++) {
-            #pragma unroll(2)
-            for (uint32_t i01 = ir0; i01 < ir1; i01++) {
-                uint8_t* dst_ptr  = (uint8_t*) dst->data  + i01*nb1  + i02*nb2  + i03*nb3;
-                uint8_t* src0_ptr = (uint8_t*) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
-                hex_l2fetch(src0_ptr, ne00 * ct->src0_type_size, nb01, 2);
-                hvx_copy_uu(dst_ptr, src0_ptr, ne00, ct->src0_type_size);
-            }
-        }
-    }
+#define DEFINE_CPY_SAMESHAPE(NAME, ELEM_TYPE, ELEM_SIZE)                                                       \
+static void cpy_thread_##NAME##_sameshape(unsigned int nth, unsigned int ith, void * data) {                   \
+    struct htp_copy_context * ct = (struct htp_copy_context *) data;                                           \
+    struct htp_ops_context * octx = ct->octx;                                                                  \
+    cpy_preamble;                                                                                              \
+    const uint32_t dr  = ct->src0_nrows_per_thread;                                                            \
+    const uint32_t ir0 = dr * ith;                                                                             \
+    const uint32_t ir1 = (ir0 + dr) < nr ? (ir0 + dr) : nr;                                                    \
+    if (ir0 >= nr) return;                                                                                     \
+    for (uint32_t i03 = 0; i03 < ne03; i03++) {                                                                \
+        for (uint32_t i02 = 0; i02 < ne02; i02++) {                                                            \
+            _Pragma("unroll(4)")                                                                               \
+            for (uint32_t i01 = ir0; i01 < ir1; i01++) {                                                       \
+                uint8_t* dst_ptr  = (uint8_t*) dst->data  + i01*nb1  + i02*nb2  + i03*nb3;                     \
+                uint8_t* src0_ptr = (uint8_t*) src0->data + i01*nb01 + i02*nb02 + i03*nb03;                    \
+                hex_l2fetch(src0_ptr, ne00 * ELEM_SIZE, nb01, 2);                                              \
+                hvx_copy_uu(dst_ptr, src0_ptr, ne00, ELEM_SIZE);                                               \
+            }                                                                                                  \
+        }                                                                                                      \
+    }                                                                                                          \
 }
 
-static void cpy_thread_sametype_reshape(struct htp_copy_context * ct, struct htp_ops_context * octx, int nth, int ith) {
-    cpy_preamble;
+DEFINE_CPY_SAMESHAPE(f32, float, 4)
+DEFINE_CPY_SAMESHAPE(f16, __fp16, 2)
 
-    // parallelize by src0 rows
-    const uint32_t dr  = ct->src0_nrows_per_thread;
-    const uint32_t ir0 = dr * ith;
-    const uint32_t ir1 = (ir0 + dr) < nr ? (ir0 + dr) : nr;
-
-    // Fast path: when both src0 and dst are contiguous in memory
-    // Replace the element-by-element loop with a single bulk HVX copy per (i03, i02) slice.
-    const bool src0_contig = (nb00 == ct->src0_type_size) &&
-                             (nb01 == ne00 * nb00) &&
-                             (nb02 == ne01 * nb01) &&
-                             (nb03 == ne02 * nb02);
-    const bool dst_contig  = (nb0  == ct->dst_type_size)  &&
-                             (nb1  == ne0  * nb0)  &&
-                             (nb2  == ne1  * nb1)  &&
-                             (nb3  == ne2  * nb2);
-
-    if (src0_contig && dst_contig) {
-        for (int64_t i03 = 0; i03 < ne03; i03++) {
-            for (int64_t i02 = 0; i02 < ne02; i02++) {
-                uint8_t * src_ptr = (uint8_t *) src0->data + i03*nb03 + i02*nb02 + ir0*nb01;
-                uint32_t  flat    = ((i03*ne02 + i02)*ne01 + ir0) * ne00;
-                uint8_t * dst_ptr = (uint8_t *) dst->data  + flat * ct->src0_type_size;
-                hvx_copy_uu(dst_ptr, src_ptr, (ir1 - ir0) * ne00, ct->src0_type_size);
-            }
-        }
-        return;
-    }
-
-    // dst counters
-    int64_t k10 = 0;
-    int64_t i11 = 0;
-    int64_t i12 = 0;
-    int64_t i13 = 0;
-
-    // number of blocks in a row
-    const int64_t nk00 = ct->src0_blocks_per_row;
-    const int64_t nk0  = ct->dst_blocks_per_row;
-
-    for (int64_t i03 = 0; i03 < ne03; i03++) {
-        for (int64_t i02 = 0; i02 < ne02; i02++) {
-            k10 += nk00 * ir0;
-            while (k10 >= nk0) {
-                k10 -= nk0;
-                if (++i11 == ne1) {
-                    i11 = 0;
-                    if (++i12 == ne2) {
-                        i12 = 0;
-                        if (++i13 == ne3) {
-                            i13 = 0;
-                        }
-                    }
-                }
-            }
-            for (int64_t i01 = ir0; i01 < ir1; i01++) {
-                for (int64_t k00 = 0; k00 < nk00; k00++) {
-                    const char * src0_ptr = ((char *) src0->data + k00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
-                          char * dst_ptr  = ((char *)  dst->data + k10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
-                    memcpy(dst_ptr, src0_ptr, ct->dst_type_size);
-
-                    if (++k10 == nk0) {
-                        k10 = 0;
-                        if (++i11 == ne1) {
-                            i11 = 0;
-                            if (++i12 == ne2) {
-                                i12 = 0;
-                                if (++i13 == ne3) {
-                                    i13 = 0;
-                                }
-                            }
-                        }
-                    }
-                }
-            }
-            k10 += nk00 * (ne01 - ir1);
-            while (k10 >= nk0) {
-                k10 -= nk0;
-                if (++i11 == ne1) {
-                    i11 = 0;
-                    if (++i12 == ne2) {
-                        i12 = 0;
-                        if (++i13 == ne3) {
-                            i13 = 0;
-                        }
-                    }
-                }
-            }
-        }
-    }
+#define DEFINE_CPY_RESHAPE(NAME, ELEM_TYPE, ELEM_SIZE)                                                         \
+static void cpy_thread_##NAME##_reshape(unsigned int nth, unsigned int ith, void * data) {                     \
+    struct htp_copy_context * ct = (struct htp_copy_context *) data;                                           \
+    struct htp_ops_context * octx = ct->octx;                                                                  \
+    cpy_preamble;                                                                                              \
+    const uint32_t dr  = ct->src0_nrows_per_thread;                                                            \
+    const uint32_t ir0 = dr * ith;                                                                             \
+    const uint32_t ir1 = (ir0 + dr) < nr ? (ir0 + dr) : nr;                                                    \
+    if (ir0 >= nr) return;                                                                                     \
+    const bool src0_contig = (nb00 == ELEM_SIZE)   &&                                                          \
+                             (nb01 == ne00 * nb00) &&                                                          \
+                             (nb02 == ne01 * nb01) &&                                                          \
+                             (nb03 == ne02 * nb02);                                                            \
+    const bool dst_contig  = (nb0  == ELEM_SIZE)   &&                                                          \
+                             (nb1  == ne0  * nb0)  &&                                                          \
+                             (nb2  == ne1  * nb1)  &&                                                          \
+                             (nb3  == ne2  * nb2);                                                             \
+    if (src0_contig && dst_contig) {                                                                           \
+        for (int64_t i03 = 0; i03 < ne03; i03++) {                                                             \
+            for (int64_t i02 = 0; i02 < ne02; i02++) {                                                         \
+                uint8_t * src_ptr = (uint8_t *) src0->data + i03*nb03 + i02*nb02 + ir0*nb01;                   \
+                uint32_t  flat    = ((i03*ne02 + i02)*ne01 + ir0) * ne00;                                      \
+                uint8_t * dst_ptr = (uint8_t *) dst->data  + flat * ELEM_SIZE;                                 \
+                hvx_copy_uu(dst_ptr, src_ptr, (ir1 - ir0) * ne00, ELEM_SIZE);                                  \
+            }                                                                                                  \
+        }                                                                                                      \
+        return;                                                                                                \
+    }                                                                                                          \
+    const bool reshape_flat_fast = (ne03 == 1 && ne2 == 1 && ne3 == 1) &&                                      \
+                                   (ne0 == ne00 * ne01) && (ne1 == ne02) &&                                    \
+                                   (nb00 == ELEM_SIZE) && (nb0 == ELEM_SIZE);                                  \
+    if (reshape_flat_fast) {                                                                                   \
+        for (uint32_t i02 = 0; i02 < ne02; i02++) {                                                            \
+            for (uint32_t i01 = ir0; i01 < ir1; i01++) {                                                       \
+                uint8_t * src0_ptr = (uint8_t *) src0->data + i01 * nb01 + i02 * nb02;                         \
+                uint8_t * dst_ptr  = (uint8_t *) dst->data  + i01 * ne00 * ELEM_SIZE + i02 * nb1;              \
+                hvx_copy_uu(dst_ptr, src0_ptr, ne00, ELEM_SIZE);                                               \
+            }                                                                                                  \
+        }                                                                                                      \
+        return;                                                                                                \
+    }                                                                                                          \
+    int64_t k10 = 0;                                                                                           \
+    int64_t i11 = 0;                                                                                           \
+    int64_t i12 = 0;                                                                                           \
+    int64_t i13 = 0;                                                                                           \
+    const int64_t nk00 = ct->src0_blocks_per_row;                                                              \
+    const int64_t nk0  = ct->dst_blocks_per_row;                                                               \
+    for (int64_t i03 = 0; i03 < ne03; i03++) {                                                                 \
+        for (int64_t i02 = 0; i02 < ne02; i02++) {                                                             \
+            k10 += nk00 * ir0;                                                                                 \
+            while (k10 >= nk0) {                                                                               \
+                k10 -= nk0;                                                                                    \
+                if (++i11 == ne1) {                                                                            \
+                    i11 = 0;                                                                                   \
+                    if (++i12 == ne2) {                                                                        \
+                        i12 = 0;                                                                               \
+                        if (++i13 == ne3) {                                                                    \
+                            i13 = 0;                                                                           \
+                        }                                                                                      \
+                    }                                                                                          \
+                }                                                                                              \
+            }                                                                                                  \
+            for (int64_t i01 = ir0; i01 < ir1; i01++) {                                                        \
+                for (int64_t k00 = 0; k00 < nk00; k00++) {                                                     \
+                    const char * src0_ptr = ((char *) src0->data + k00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); \
+                          char * dst_ptr  = ((char *)  dst->data + k10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);  \
+                    memcpy(dst_ptr, src0_ptr, ELEM_SIZE);                                                      \
+                    if (++k10 == nk0) {                                                                        \
+                        k10 = 0;                                                                               \
+                        if (++i11 == ne1) {                                                                    \
+                            i11 = 0;                                                                           \
+                            if (++i12 == ne2) {                                                                \
+                                i12 = 0;                                                                       \
+                                if (++i13 == ne3) {                                                            \
+                                    i13 = 0;                                                                   \
+                                }                                                                              \
+                            }                                                                                  \
+                        }                                                                                      \
+                    }                                                                                          \
+                }                                                                                              \
+            }                                                                                                  \
+            k10 += nk00 * (ne01 - ir1);                                                                        \
+            while (k10 >= nk0) {                                                                               \
+                k10 -= nk0;                                                                                    \
+                if (++i11 == ne1) {                                                                            \
+                    i11 = 0;                                                                                   \
+                    if (++i12 == ne2) {                                                                        \
+                        i12 = 0;                                                                               \
+                        if (++i13 == ne3) {                                                                    \
+                            i13 = 0;                                                                           \
+                        }                                                                                      \
+                    }                                                                                          \
+                }                                                                                              \
+            }                                                                                                  \
+        }                                                                                                      \
+    }                                                                                                          \
 }
 
-static void cpy_thread_f16_f32_sameshape(struct htp_copy_context * ct, struct htp_ops_context * octx, const int nth, const int ith) {
+DEFINE_CPY_RESHAPE(f32, float, 4)
+DEFINE_CPY_RESHAPE(f16, __fp16, 2)
+
+static void cpy_thread_f16_f32_sameshape(unsigned int nth, unsigned int ith, void * data) {
+    struct htp_copy_context * ct = (struct htp_copy_context *) data;
+    struct htp_ops_context * octx = ct->octx;
     cpy_preamble;
 
     // parallelize by src0 rows
     const uint32_t dr  = ct->src0_nrows_per_thread;
     const uint32_t ir0 = dr * ith;
     const uint32_t ir1 = (ir0 + dr) < nr ? (ir0 + dr) : nr;
+    if (ir0 >= nr) return;
 
     // copy by rows
     for (uint32_t i03 = 0; i03 < ne03; i03++) {
@@ -195,13 +207,16 @@ static void cpy_thread_f16_f32_sameshape(struct htp_copy_context * ct, struct ht
     }
 }
 
-static void cpy_thread_f32_f16_sameshape(struct htp_copy_context * ct, struct htp_ops_context * octx, const int nth, const int ith) {
+static void cpy_thread_f32_f16_sameshape(unsigned int nth, unsigned int ith, void * data) {
+    struct htp_copy_context * ct = (struct htp_copy_context *) data;
+    struct htp_ops_context * octx = ct->octx;
     cpy_preamble;
 
     // parallelize by src0 rows
     const uint32_t dr  = ct->src0_nrows_per_thread;
     const uint32_t ir0 = dr * ith;
     const uint32_t ir1 = (ir0 + dr) < nr ? (ir0 + dr) : nr;
+    if (ir0 >= nr) return;
 
     // copy by rows
     for (uint32_t i03 = 0; i03 < ne03; i03++) {
@@ -217,11 +232,6 @@ static void cpy_thread_f32_f16_sameshape(struct htp_copy_context * ct, struct ht
     }
 }
 
-static void cpy_work_func(unsigned int n, unsigned int i, void *data) {
-    struct htp_copy_context *ct = (struct htp_copy_context *) data;
-    ct->copy(ct, ct->octx, n, i);
-}
-
 int op_cpy(struct htp_ops_context * octx) {
     cpy_preamble;
 
@@ -254,22 +264,32 @@ int op_cpy(struct htp_ops_context * octx) {
 
     ct.src0_nrows_per_thread = (nr + n_threads - 1) / n_threads;
 
+    worker_callback_t copy_fun;
+
     if (sametype && sameshape) {
-        ct.copy = cpy_thread_sametype_sameshape;
+        if (src0->type == HTP_TYPE_F32) {
+            copy_fun = cpy_thread_f32_sameshape;
+        } else {
+            copy_fun = cpy_thread_f16_sameshape;
+        }
     } else if (sameshape) {
         /**/ if (dst->type == HTP_TYPE_F16 && src0->type == HTP_TYPE_F32)
-            ct.copy = cpy_thread_f16_f32_sameshape;
+            copy_fun = cpy_thread_f16_f32_sameshape;
         else if (dst->type == HTP_TYPE_F32 && src0->type == HTP_TYPE_F16)
-            ct.copy = cpy_thread_f32_f16_sameshape;
+            copy_fun = cpy_thread_f32_f16_sameshape;
         else
             return HTP_STATUS_NO_SUPPORT;
     } else if (sametype) {
-        ct.copy = cpy_thread_sametype_reshape;
+        if (src0->type == HTP_TYPE_F32) {
+            copy_fun = cpy_thread_f32_reshape;
+        } else {
+            copy_fun = cpy_thread_f16_reshape;
+        }
     } else {
         return HTP_STATUS_NO_SUPPORT;
     }
 
-    worker_pool_run_func(octx->ctx->worker_pool, cpy_work_func, &ct, n_threads);
+    worker_pool_run_func(octx->ctx->worker_pool, copy_fun, &ct, n_threads);
 
     return HTP_STATUS_OK;
 }

ggml/src/ggml-hexagon/htp/flash-attn-ops.c

@@ -22,6 +22,16 @@
 // Must be multiple of 32
 #define FLASH_ATTN_BLOCK_SIZE (32 * 2)
 
+#if __HVX_ARCH__ < 79
+#define HVX_OP_ADD_F32(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(a, b))
+#define HVX_OP_SUB_F32(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vsub_VsfVsf(a, b))
+#define HVX_OP_MUL_F32(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(a, b))
+#else
+#define HVX_OP_ADD_F32(a, b) Q6_Vsf_vadd_VsfVsf(a, b)
+#define HVX_OP_SUB_F32(a, b) Q6_Vsf_vsub_VsfVsf(a, b)
+#define HVX_OP_MUL_F32(a, b) Q6_Vsf_vmpy_VsfVsf(a, b)
+#endif
+
 // This is a bit of a hack because the compiler is strugling to properly inline
 // the default hvx_vec_f32_to_f16 with output into the local array.
 static __attribute__((noinline)) void hvx_vec_f32_to_f16_a(void *ptr, HVX_Vector v0, HVX_Vector v1)
@@ -54,8 +64,8 @@ static inline void hvx_dot_f16_f16_aa(float * restrict r, const void * restrict
         rsum_p = hvx_vec_mpyacc_f32_f16(rsum_p, x_hf, y_hf);
     }
 
-    HVX_Vector rsum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(rsum_p), Q6_V_hi_W(rsum_p)));
-    rsum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(hvx_vec_splat_f32(s), hvx_vec_reduce_sum_f32(rsum)));
+    HVX_Vector rsum = HVX_OP_ADD_F32(Q6_V_lo_W(rsum_p), Q6_V_hi_W(rsum_p));
+    rsum = HVX_OP_MUL_F32(hvx_vec_splat_f32(s), hvx_vec_reduce_sum_f32(rsum));
     hvx_vec_store_u(r, 4, rsum);
 }
 
@@ -105,10 +115,10 @@ static inline HVX_Vector hvx_dot_f16_f16_aa_rx4(const void * restrict y,
         rsum3_p = hvx_vec_mpyacc_f32_f16(rsum3_p, x3_hf, y_hf);
     }
 
-    HVX_Vector rsum0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(rsum0_p), Q6_V_hi_W(rsum0_p)));
-    HVX_Vector rsum1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(rsum1_p), Q6_V_hi_W(rsum1_p)));
-    HVX_Vector rsum2 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(rsum2_p), Q6_V_hi_W(rsum2_p)));
-    HVX_Vector rsum3 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(rsum3_p), Q6_V_hi_W(rsum3_p)));
+    HVX_Vector rsum0 = HVX_OP_ADD_F32(Q6_V_lo_W(rsum0_p), Q6_V_hi_W(rsum0_p));
+    HVX_Vector rsum1 = HVX_OP_ADD_F32(Q6_V_lo_W(rsum1_p), Q6_V_hi_W(rsum1_p));
+    HVX_Vector rsum2 = HVX_OP_ADD_F32(Q6_V_lo_W(rsum2_p), Q6_V_hi_W(rsum2_p));
+    HVX_Vector rsum3 = HVX_OP_ADD_F32(Q6_V_lo_W(rsum3_p), Q6_V_hi_W(rsum3_p));
 
     HVX_Vector_x4 rsum0123 = { .v = { rsum0, rsum1, rsum2, rsum3 } };
     return hvx_vec_reduce_sum_f32x4(rsum0123);
@@ -123,7 +133,7 @@ static inline HVX_Vector hvx_dot_f16_f16_aa_rx32(const void * restrict y,
     const size_t nvec = n / VLEN_FP16; // num full fp16 hvx vectors
     const size_t nloe = n % VLEN_FP16; // leftover elements
 
-    HVX_Vector   sums;  // initialize at j = 0
+    HVX_Vector   sums = Q6_V_vzero();
     const size_t stride_x_4 = stride_x * 4;
     for (uint32_t j = 0; j < VLEN_FP32; j += 4) {
         HVX_Vector     sums_x4 = hvx_dot_f16_f16_aa_rx4(y, x, stride_x, nvec, nloe);
@@ -132,8 +142,7 @@ static inline HVX_Vector hvx_dot_f16_f16_aa_rx32(const void * restrict y,
         x += stride_x_4;
     }
 
-    sums = Q6_Vqf32_vmpy_VsfVsf(hvx_vec_splat_f32(s), sums);
-    return Q6_Vsf_equals_Vqf32(sums);
+    return HVX_OP_MUL_F32(hvx_vec_splat_f32(s), sums);
 }
 
 // MAD: y (F32) += x (F16) * s (F16)
@@ -268,11 +277,10 @@ static inline void hvx_scale_vec_f32_aa(uint8_t * restrict dst, const uint8_t *
     uint32_t i = 0;
     #pragma unroll(4)
     for (; i < nvec; ++i) {
-        vdst[i] = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(vsrc[i], vs));
+        vdst[i] = HVX_OP_MUL_F32(vsrc[i], vs);
     }
     if (nloe) {
-        HVX_Vector v = Q6_Vqf32_vmpy_VsfVsf(vsrc[i], vs);
-        hvx_vec_store_a(&vdst[i], nloe * sizeof(float), Q6_Vsf_equals_Vqf32(v));
+        hvx_vec_store_a(&vdst[i], nloe * sizeof(float), HVX_OP_MUL_F32(vsrc[i], vs));
     }
 }
 
@@ -438,25 +446,44 @@ static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void *
             // Process in sub-blocks of 32 (VLEN_FP32)
             HVX_Vector sb_scores[FLASH_ATTN_BLOCK_SIZE / VLEN_FP32];
             HVX_Vector v_max = hvx_vec_splat_f32(-INFINITY);
-            for (uint32_t iv = 0; ic + VLEN_FP32 <= current_block_size; ic += VLEN_FP32, ++iv) {
+            for (uint32_t iv = 0; ic < current_block_size; ic += VLEN_FP32, ++iv) {
                 // 1. Compute scores
                 HVX_Vector scores = hvx_dot_f16_f16_aa_rx32(q_ptr_vtcm, k_base + ic * factx->size_k_row_padded, factx->size_k_row_padded, DK, factx->scale);
 
                 // 2. Softcap
                 if (factx->logit_softcap != 0.0f) {
                     scores = hvx_vec_tanh_f32(scores);
-                    scores = Q6_Vqf32_vmpy_VsfVsf(scores, logit_cap);
-                    scores = Q6_Vsf_equals_Vqf32(scores);
+                    scores = HVX_OP_MUL_F32(scores, logit_cap);
                 }
 
                 // 3. Mask
                 if (mask) {
                     const __fp16 * mp = m_base + ic;
                     HVX_Vector m_vals_f16 = *(const HVX_UVector *) mp;
-                    HVX_VectorPair m_vals_f32_pair = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(m_vals_f16), slope_vec);
-                    HVX_Vector add_val = Q6_V_lo_W(m_vals_f32_pair);
-                    scores = Q6_Vqf32_vadd_Vqf32Vsf(add_val, scores);
-                    scores = Q6_Vsf_equals_Vqf32(scores);
+
+                    // Multiplying -INFINITY (0xFC00) by a slope in VhfVhf instructions can incorrectly produce NaN on v79.
+                    // Clamp -INFINITY to the max negative fp16 finite value (-65504.0f).
+                    HVX_Vector vinf = Q6_Vh_vsplat_R(0xFC00);
+                    HVX_Vector vmin = Q6_Vh_vsplat_R(0xFBFF);
+                    HVX_VectorPred is_inf = Q6_Q_vcmp_eq_VhVh(m_vals_f16, vinf);
+                    m_vals_f16 = Q6_V_vmux_QVV(is_inf, vmin, m_vals_f16);
+
+                    #if __HVX_ARCH__ >= 79
+                        HVX_VectorPair m_vals_f32_pair = Q6_Wsf_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(m_vals_f16), slope_vec);
+                        HVX_Vector add_val = Q6_V_lo_W(m_vals_f32_pair);
+                        scores = Q6_Vsf_vadd_VsfVsf(add_val, scores);
+                    #else
+                        HVX_VectorPair m_vals_f32_pair = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(m_vals_f16), slope_vec);
+                        HVX_Vector add_val = Q6_V_lo_W(m_vals_f32_pair);
+                        scores = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(add_val, scores));
+                    #endif
+                }
+
+                // Mask out invalid lanes for leftover handling
+                uint32_t valid_lanes = current_block_size - ic;
+                if (valid_lanes < VLEN_FP32) {
+                    HVX_VectorPred valid_pred = Q6_Q_vsetq_R(valid_lanes * 4); // 4 bytes per fp32 lane
+                    scores = Q6_V_vmux_QVV(valid_pred, scores, hvx_vec_splat_f32(-INFINITY));
                 }
 
                 sb_scores[iv] = scores;
@@ -466,78 +493,55 @@ static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void *
             {
                 // 4. Online Softmax Update
                 HVX_Vector M_new_vec = Q6_Vsf_vmax_VsfVsf(v_max, M_vec);
-                HVX_Vector diff_vec  = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vsub_VsfVsf(M_vec, M_new_vec));
+                HVX_Vector diff_vec  = HVX_OP_SUB_F32(M_vec, M_new_vec);
                 HVX_Vector ms_vec    = hvx_vec_exp_f32(diff_vec);
                 M_vec = M_new_vec;
 
                 hvx_scale_vec_f32_aa((uint8_t *) VKQ32, (const uint8_t *) VKQ32, DV, ms_vec);
 
                 HVX_Vector p_sum_vec = hvx_vec_splat_f32(0.0f);
-                for (uint32_t ic2 = 0, iv = 0; ic2 + VLEN_FP32 <= current_block_size; ic2 += VLEN_FP32, ++iv) {
+                for (uint32_t ic2 = 0, iv = 0; ic2 < current_block_size; ic2 += VLEN_FP32, ++iv) {
                     HVX_Vector scores = sb_scores[iv];
-                    HVX_Vector scores_shifted = Q6_Vqf32_vsub_VsfVsf(scores, M_vec);
-                    HVX_Vector P = hvx_vec_exp_f32(Q6_Vsf_equals_Vqf32(scores_shifted));
+                    HVX_Vector scores_shifted = HVX_OP_SUB_F32(scores, M_vec);
+                    HVX_Vector P = hvx_vec_exp_f32(scores_shifted);
 
-                    p_sum_vec = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(p_sum_vec, P));
+                    p_sum_vec = HVX_OP_ADD_F32(p_sum_vec, P);
 
                     // 5. Accumulate V
                     __fp16 __attribute__((aligned(VLEN))) p_arr[VLEN_FP16];
                     hvx_vec_f32_to_f16_a(p_arr, P, hvx_vec_splat_f32(0));
 
+                    float __attribute__((aligned(128))) P_arr[VLEN_FP32];
+                    hvx_vec_store_a(P_arr, 128, P);
+
                     for (uint32_t j = 0; j < VLEN_FP32; j += 2) {
-                        const uint32_t  cur_ic = ic2 + j;
-                        const uint8_t * v_ptr  = v_base + cur_ic * factx->size_v_row_padded;
+                        const uint32_t cur_ic = ic2 + j;
+                        if (cur_ic >= current_block_size) {
+                            break;
+                        }
+
+                        if (cur_ic + 1 == current_block_size) {
+                            // Odd leftover, process single row
+                            if (P_arr[j] != 0.0f) {
+                                const uint8_t * v_ptr = v_base + cur_ic * factx->size_v_row_padded;
+                                hvx_mad_f32_f16_aa(VKQ32, v_ptr, (p_arr + j), DV);
+                            }
+                            break;
+                        }
+
+                        // Avoid NaN * 0.0 = NaN for uninitialized V cache rows.
+                        // Check the f32 values to safely avoid strict aliasing violations.
+                        if (P_arr[j] == 0.0f && P_arr[j + 1] == 0.0f) {
+                            continue;
+                        }
+
+                        const uint8_t * v_ptr = v_base + cur_ic * factx->size_v_row_padded;
                         hvx_mad_f32_f16_aa_rx2(VKQ32, v_ptr, v_ptr + factx->size_v_row_padded, (p_arr + j), (p_arr + j + 1), DV);
                     }
                 }
 
                 p_sum_vec = hvx_vec_reduce_sum_f32(p_sum_vec);
-                S_vec = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(S_vec, ms_vec)), p_sum_vec));
-            }
-
-            if (ic < current_block_size) {
-                // Sync scalars for leftover/next block if needed
-                float M = hvx_vec_get_f32(M_vec);
-                float S = hvx_vec_get_f32(S_vec);
-
-                // Leftover
-                for (; ic < current_block_size; ++ic) {
-                    float s_val;
-                    const uint8_t * k_ptr = k_base + ic * factx->size_k_row_padded;
-                    hvx_dot_f16_f16_aa(&s_val, q_ptr_vtcm, k_ptr, DK, factx->scale);
-                    if (factx->logit_softcap != 0.0f) {
-                        s_val = factx->logit_softcap * tanhf(s_val);
-                    }
-
-                    if (mask) {
-                        const float m_val = m_base[ic];
-                        s_val += slope * m_val;
-                    }
-
-                    const float Mold = M;
-                    __fp16 vs = 1.0f;
-
-                    if (s_val > M) {
-                        M = s_val;
-                        HVX_Vector diff_vec = hvx_vec_splat_f32(Mold - M);
-                        HVX_Vector ms_vec   = hvx_vec_exp_f32(diff_vec);
-                        hvx_scale_vec_f32_aa((uint8_t *) VKQ32, (const uint8_t *) VKQ32, DV, ms_vec);
-
-                        float ms = hvx_vec_get_f32(ms_vec);
-                        S = S * ms + vs;
-                    } else {
-                        HVX_Vector diff_vec = hvx_vec_splat_f32(s_val - M);
-                        vs = hvx_vec_get_f32(hvx_vec_exp_f32(diff_vec));
-                        S += vs;
-                    }
-
-                    const uint8_t * v_ptr = v_base + ic * factx->size_v_row_padded;
-
-                    hvx_mad_f32_f16_aa(VKQ32, v_ptr, &vs, DV);
-                }
-
-                M_vec = hvx_vec_splat_f32(M);
-                S_vec = hvx_vec_splat_f32(S);
+                S_vec = HVX_OP_ADD_F32(HVX_OP_MUL_F32(S_vec, ms_vec), p_sum_vec);
             }
 
             // Issue DMA for next+1 block (if exists)
@@ -599,8 +603,9 @@ static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void *
         const int i2 = iq2;
         const int i3 = iq3;
 
-        // dst is permuted
-        uint8_t * dst_ptr = (uint8_t *) dst->data + (i3*ne2*ne1 + i2 + i1*ne1) * nb1;
+        // dst is permuted: [DV, n_heads, n_tokens, n_seq]
+        // head stride is nb[1], token stride is nb[2], batch stride is nb[3]
+        uint8_t * dst_ptr = (uint8_t *) dst->data + i2 * dst->nb[1] + i1 * dst->nb[2] + i3 * dst->nb[3];
 
         if (dst->type == HTP_TYPE_F32) {
             hvx_copy_f32_ua(dst_ptr, (uint8_t *) VKQ32, DV);
@@ -623,8 +628,8 @@ int op_flash_attn_ext(struct htp_ops_context * octx) {
     }
 
 #ifdef HTP_HAS_HMX
-    // HMX path: prefill (neq1 >= 32), head_dim multiple of 32, F16 KV
-    if (k->type == HTP_TYPE_F16 && v->type == HTP_TYPE_F16 && k->ne[0] % 32 == 0 && q->ne[1] >= 32) {
+    // HMX path: head_dim multiple of 32, F16 KV
+    if (k->type == HTP_TYPE_F16 && v->type == HTP_TYPE_F16 && k->ne[0] % 32 == 0) {
         int ret = hmx_flash_attn_ext(octx);
         if (ret == HTP_STATUS_OK) {
             return ret;

ggml/src/ggml-hexagon/htp/gated-delta-net-ops.c

@@ -586,6 +586,7 @@ static void gated_delta_net_f32_pp_thread(unsigned int nth, unsigned int ith, vo
     const uint32_t H        = v->ne[1];
     const uint32_t n_tokens = v->ne[2];
     const uint32_t n_seqs   = v->ne[3];
+    const uint32_t K        = state->ne[1];
 
     const uint32_t total_rows = H * n_seqs;
     if (ith >= total_rows) {
@@ -606,6 +607,10 @@ static void gated_delta_net_f32_pp_thread(unsigned int nth, unsigned int ith, vo
     float local_k[HTP_GDN_MAX_SV] __attribute__((aligned(128)));
     float local_sums[4] __attribute__((aligned(128)));
 
+    const uint64_t state_seq_stride = state->nb[2] / sizeof(float);
+    const uint64_t state_size_per_snap = (uint64_t) S_v * S_v * H * n_seqs;
+    const int64_t shift = (int64_t) n_tokens - (int64_t) K;
+
     for (uint32_t ir = ith; ir < total_rows; ir += nth) {
         const uint32_t iv1 = ir % H;
         const uint32_t iv3 = ir / H;
@@ -615,8 +620,8 @@ static void gated_delta_net_f32_pp_thread(unsigned int nth, unsigned int ith, vo
         const uint32_t iq3 = iv3 / rq3;
         const uint32_t ik3 = iv3 / rk3;
 
-        float * s_out = state_out_base + ((uint64_t) iv3 * H + iv1) * S_v * S_v;
-        const float * s_in = state_in_base + ((uint64_t) iv3 * H + iv1) * S_v * S_v;
+        float * s_out = state_out_base + (uint64_t) (K - 1) * state_size_per_snap + ((uint64_t) iv3 * H + iv1) * S_v * S_v;
+        const float * s_in = state_in_base + (uint64_t) iv3 * state_seq_stride + (uint64_t) iv1 * S_v * S_v;
 
         memcpy(s_out, s_in, gctx->state_bytes);
         float * s_work = s_out;
@@ -689,6 +694,16 @@ static void gated_delta_net_f32_pp_thread(unsigned int nth, unsigned int ith, vo
                 }
             }
 
+            if (K > 1) {
+                const int64_t target_slot = (int64_t) t - shift;
+                if (target_slot >= 0 && target_slot < (int64_t) K) {
+                    float * curr_state_o = state_out_base + (uint64_t) target_slot * state_size_per_snap + ((uint64_t) iv3 * H + iv1) * S_v * S_v;
+                    if (curr_state_o != s_work) {
+                        memcpy(curr_state_o, s_work, gctx->state_bytes);
+                    }
+                }
+            }
+
             attn_data += (uint64_t) S_v * H;
         }
     }
@@ -709,6 +724,7 @@ static void gated_delta_net_f32_tg_thread(unsigned int nth, unsigned int ith, vo
     const uint32_t S_v      = v->ne[0];
     const uint32_t H        = v->ne[1];
     const uint32_t n_seqs   = v->ne[3];
+    const uint32_t K        = state->ne[1];
 
     const uint32_t total_rows = H * n_seqs;
     if (ith >= total_rows) {
@@ -736,6 +752,9 @@ static void gated_delta_net_f32_tg_thread(unsigned int nth, unsigned int ith, vo
         spad = gctx->vtcm_state_base + gctx->vtcm_state_per_thread * ith;
     }
 
+    const uint64_t state_seq_stride = state->nb[2] / sizeof(float);
+    const uint64_t state_size_per_snap = (uint64_t) S_v * S_v * H * n_seqs;
+
     for (uint32_t ir = ith; ir < total_rows; ir += nth) {
         const uint32_t iv1 = ir % H;
         const uint32_t iv3 = ir / H;
@@ -745,8 +764,8 @@ static void gated_delta_net_f32_tg_thread(unsigned int nth, unsigned int ith, vo
         const uint32_t iq3 = iv3 / rq3;
         const uint32_t ik3 = iv3 / rk3;
 
-        float * s_out = state_out_base + ((uint64_t) iv3 * H + iv1) * S_v * S_v;
-        const float * s_in = state_in_base + ((uint64_t) iv3 * H + iv1) * S_v * S_v;
+        float * s_out = state_out_base + (uint64_t) (K - 1) * state_size_per_snap + ((uint64_t) iv3 * H + iv1) * S_v * S_v;
+        const float * s_in = state_in_base + (uint64_t) iv3 * state_seq_stride + (uint64_t) iv1 * S_v * S_v;
         float * s_work;
 
         if (spad) {
@@ -901,6 +920,7 @@ int op_gated_delta_net(struct htp_ops_context * octx) {
     const uint32_t H        = v->ne[1];
     const uint32_t n_tokens = v->ne[2];
     const uint32_t n_seqs   = v->ne[3];
+    const uint32_t K        = state->ne[1];
 
     if (S_v == 0 || S_v > HTP_GDN_MAX_SV || H == 0 || n_tokens == 0 || n_seqs == 0) {
         return HTP_STATUS_NO_SUPPORT;
@@ -913,10 +933,10 @@ int op_gated_delta_net(struct htp_ops_context * octx) {
         (n_seqs % q->ne[3]) != 0 || (n_seqs % k->ne[3]) != 0) {
         return HTP_STATUS_NO_SUPPORT;
     }
-    if (state->ne[0] * state->ne[1] * state->ne[2] * state->ne[3] != S_v * S_v * H * n_seqs) {
+    if (state->ne[0] * state->ne[2] * state->ne[3] != S_v * S_v * H * n_seqs) {
         return HTP_STATUS_NO_SUPPORT;
     }
-    if (dst->ne[0] != S_v * H || dst->ne[1] != n_tokens * n_seqs + S_v * n_seqs) {
+    if (dst->ne[0] != S_v * H || dst->ne[1] != n_tokens * n_seqs + S_v * n_seqs * K) {
         return HTP_STATUS_NO_SUPPORT;
     }
 

ggml/src/ggml-hexagon/htp/get-rows-ops.c

@@ -17,9 +17,13 @@
 
 struct get_rows_context {
     struct htp_ops_context * octx;
-    uint32_t src1_nrows_per_thread;
+    uint32_t tasks_per_thread;
+    uint32_t total_tasks;
+    uint32_t chunks_per_row;
+    uint32_t chunk_size;
     struct fastdiv_values get_rows_div_ne10;
     struct fastdiv_values get_rows_div_ne10_ne11;
+    struct fastdiv_values get_rows_div_chunks_per_row;
 };
 
 #define get_rows_preamble \
@@ -52,20 +56,23 @@ struct get_rows_context {
                                                \
     const uint32_t nr = ne10 * ne11 * ne12;
 
-static void get_rows_thread_f32_f32(unsigned int nth, unsigned int ith, void *data) {
+static void get_rows_thread_f32_f32_dma(unsigned int nth, unsigned int ith, void *data) {
     struct get_rows_context * grctx = (struct get_rows_context *)data;
     struct htp_ops_context * octx = grctx->octx;
     get_rows_preamble;
 
     uint64_t qt = HAP_perf_get_qtimer_count();
 
-    // parallelize by src1 elements (which correspond to dst rows)
-    const uint32_t dr  = grctx->src1_nrows_per_thread;
+    const uint32_t dr  = grctx->tasks_per_thread;
     const uint32_t ir0 = dr * ith;
-    const uint32_t ir1 = (ir0 + dr < nr) ? (ir0 + dr) : nr;
+    if (ir0 >= grctx->total_tasks) {
+        return;
+    }
+    const uint32_t ir1 = MIN(ir0 + dr, grctx->total_tasks);
 
     const bool is_i32 = (octx->src[1]->type == HTP_TYPE_I32);
 
+    dma_queue * dma_queue = octx->ctx->dma[ith];
     for (uint32_t i = ir0; i < ir1; ++i) {
         const uint32_t i12 = fastdiv(i, &grctx->get_rows_div_ne10_ne11);
         const uint32_t rem = i - i12 * ne11 * ne10;
@@ -73,29 +80,77 @@ static void get_rows_thread_f32_f32(unsigned int nth, unsigned int ith, void *da
         const uint32_t i10 = rem - i11 * ne10;
 
         const uintptr_t src1_addr = octx->src[1]->data + i10*nb10 + i11*nb11 + i12*nb12;
-
         uint32_t i01 = is_i32 ? *(int32_t *)src1_addr : *(int64_t *)src1_addr;
 
         if (i01 >= ne01) {
-            // invalid index, skip for now to avoid crash
             continue;
         }
 
         const uintptr_t src0_ptr = octx->src[0]->data + i01*nb01 + i11*nb02 + i12*nb03;
         const uintptr_t dst_ptr  = octx->dst->data    + i10*nb1  + i11*nb2  + i12*nb3;
-        hvx_copy_f32_uu((uint8_t *)dst_ptr, (const uint8_t *)src0_ptr, ne00);
+
+        while (!dma_queue_push(dma_queue, dma_make_ptr((void *)dst_ptr, (const void *)src0_ptr), nb1, nb01, ne00 * sizeof(float), 1)) {
+            dma_queue_pop(dma_queue);
+        }
+    }
+    dma_queue_flush(dma_queue);
+
+    qt = HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - qt);
+    FARF(HIGH, "get-rows-f32-f32-dma %d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n", ith, nth,
+         ne00, ne01, ne02, ne03, ir0, ir1, ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3, (unsigned) qt);
+}
+
+static void get_rows_thread_f32_f32_hvx(unsigned int nth, unsigned int ith, void *data) {
+    struct get_rows_context * grctx = (struct get_rows_context *)data;
+    struct htp_ops_context * octx = grctx->octx;
+    get_rows_preamble;
+
+    uint64_t qt = HAP_perf_get_qtimer_count();
+
+    const uint32_t dr  = grctx->tasks_per_thread;
+    const uint32_t ir0 = dr * ith;
+    if (ir0 >= grctx->total_tasks) {
+        return;
+    }
+    const uint32_t ir1 = MIN(ir0 + dr, grctx->total_tasks);
+
+    const bool is_i32 = (octx->src[1]->type == HTP_TYPE_I32);
+
+    const uint32_t chunks_per_row = grctx->chunks_per_row;
+    const uint32_t chunk_size     = grctx->chunk_size;
+    for (uint32_t i = ir0; i < ir1; ++i) {
+        const uint32_t row_idx   = fastdiv(i, &grctx->get_rows_div_chunks_per_row);
+        const uint32_t chunk_idx = i - row_idx * chunks_per_row;
+
+        const uint32_t i12 = fastdiv(row_idx, &grctx->get_rows_div_ne10_ne11);
+        const uint32_t rem = row_idx - i12 * ne11 * ne10;
+        const uint32_t i11 = fastdiv(rem, &grctx->get_rows_div_ne10);
+        const uint32_t i10 = rem - i11 * ne10;
+
+        const uintptr_t src1_addr = octx->src[1]->data + i10*nb10 + i11*nb11 + i12*nb12;
+        uint32_t i01 = is_i32 ? *(int32_t *)src1_addr : *(int64_t *)src1_addr;
+
+        if (i01 >= ne01) {
+            continue;
+        }
+
+        const uint32_t offset = chunk_idx * chunk_size;
+        if (offset < ne00) {
+            const uint32_t copy_size = MIN(chunk_size, ne00 - offset);
+            const uintptr_t src0_ptr = octx->src[0]->data + i01*nb01 + i11*nb02 + i12*nb03 + offset * sizeof(float);
+            const uintptr_t dst_ptr  = octx->dst->data    + i10*nb1  + i11*nb2  + i12*nb3  + offset * sizeof(float);
+            hvx_copy_f32_uu((uint8_t *)dst_ptr, (const uint8_t *)src0_ptr, copy_size);
+        }
     }
 
     qt = HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - qt);
-    FARF(HIGH, "get-rows-f32-f32 %d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n", ith, nth,
+    FARF(HIGH, "get-rows-f32-f32-hvx %d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n", ith, nth,
          ne00, ne01, ne02, ne03, ir0, ir1, ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3, (unsigned) qt);
 }
 
 int op_get_rows(struct htp_ops_context * octx) {
     get_rows_preamble;
 
-    const uint32_t n_threads = MIN(nr, octx->n_threads);
-
     if (octx->src[0]->type != HTP_TYPE_F32) {
         return HTP_STATUS_NO_SUPPORT;
     }
@@ -112,13 +167,52 @@ int op_get_rows(struct htp_ops_context * octx) {
         return HTP_STATUS_OK;
     }
 
+    const uint32_t nb00 = octx->src[0]->nb[0];
+    const uint32_t nb0  = octx->dst->nb[0];
+
+    const bool can_use_dma = (nb00 == sizeof(float)) && (nb0 == sizeof(float));
+    const bool use_dma = can_use_dma && (ne00 >= 2048);
+
     struct get_rows_context grctx;
     grctx.octx = octx;
     grctx.get_rows_div_ne10      = init_fastdiv_values(octx->src[1]->ne[0]);
     grctx.get_rows_div_ne10_ne11 = init_fastdiv_values(octx->src[1]->ne[0] * octx->src[1]->ne[1]);
 
-    grctx.src1_nrows_per_thread = (nr + n_threads - 1) / n_threads;
+    if (use_dma) {
+        grctx.chunks_per_row = 1;
+        grctx.chunk_size = ne00;
+        grctx.total_tasks = nr;
+        grctx.get_rows_div_chunks_per_row = init_fastdiv_values(1);
 
-    worker_pool_run_func(octx->ctx->worker_pool, get_rows_thread_f32_f32, &grctx, n_threads);
+        const uint32_t n_threads = MIN(nr, octx->n_threads);
+        grctx.tasks_per_thread = (nr + n_threads - 1) / n_threads;
+
+        worker_pool_run_func(octx->ctx->worker_pool, get_rows_thread_f32_f32_dma, &grctx, n_threads);
+    } else {
+        uint32_t chunks_per_row = 1;
+        uint32_t chunk_size = ne00;
+        uint32_t total_tasks = nr;
+
+        if (nr < octx->n_threads) {
+            const uint32_t min_chunk_size = 1024;
+            uint32_t max_chunks = ne00 / min_chunk_size;
+            if (max_chunks == 0) {
+                max_chunks = 1;
+            }
+            chunks_per_row = MIN((octx->n_threads + nr - 1) / nr, max_chunks);
+            chunk_size = (ne00 + chunks_per_row - 1) / chunks_per_row;
+            total_tasks = nr * chunks_per_row;
+        }
+
+        grctx.chunks_per_row = chunks_per_row;
+        grctx.chunk_size = chunk_size;
+        grctx.total_tasks = total_tasks;
+        grctx.get_rows_div_chunks_per_row = init_fastdiv_values(chunks_per_row);
+
+        const uint32_t n_threads = MIN(total_tasks, octx->n_threads);
+        grctx.tasks_per_thread = (total_tasks + n_threads - 1) / n_threads;
+
+        worker_pool_run_func(octx->ctx->worker_pool, get_rows_thread_f32_f32_hvx, &grctx, n_threads);
+    }
     return HTP_STATUS_OK;
 }

ggml/src/ggml-hexagon/htp/hmx-flash-attn-ops.c

@@ -50,8 +50,8 @@ static size_t hmx_fa_compute_vtcm_usage(size_t gqa_factor, size_t DK, size_t DV,
     const size_t g_br         = hex_align_up(gqa_factor * Br, HMX_FP16_TILE_N_ROWS);
     const size_t q_tile_size  = hex_align_up(g_br * DK * sizeof(__fp16), 4096);    // Q:  [g_br, DK]
     const size_t o_tile_size  = hex_align_up(g_br * DV * sizeof(__fp16), 4096);    // O:  [g_br, DV] x2 ping-pong
-    const size_t k_dma_size   = hex_align_up(Bc * DK * sizeof(__fp16), 4096);      // K DMA: [Bc, DK] x2 double-buf
-    const size_t v_dma_size   = hex_align_up(Bc * DV * sizeof(__fp16), 4096);      // V DMA: [Bc, DV] x2 double-buf
+    const size_t k_dma_size   = hex_align_up(Bc * hex_round_up(DK * sizeof(__fp16), 128), 4096);      // K DMA: [Bc, DK] x2 double-buf
+    const size_t v_dma_size   = hex_align_up(Bc * hex_round_up(DV * sizeof(__fp16), 128), 4096);      // V DMA: [Bc, DV] x2 double-buf
     const size_t k_tile_size  = hex_align_up(Bc * DK * sizeof(__fp16), 4096);      // K tiles: [Bc, DK] interleaved
     const size_t v_tile_size  = hex_align_up(Bc * DV * sizeof(__fp16), 4096);      // V tiles: [Bc, DV] interleaved
     const size_t s_tile_size  = hex_align_up(g_br * Bc * sizeof(__fp16), 4096);    // S/P:[g_br, Bc]
@@ -1248,9 +1248,6 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
     if (DK % 32 != 0 || DV % 32 != 0) {
         return HTP_STATUS_NO_SUPPORT;
     }
-    if (neq1 < 32) {
-        return HTP_STATUS_NO_SUPPORT;
-    }
 
     // GQA factor
     const uint32_t n_kv_heads = k->ne[2];
@@ -1278,7 +1275,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
     struct hmx_fa_context factx;
     memset(&factx, 0, sizeof(factx));
     factx.octx           = octx;
-    factx.n_threads      = octx->ctx->n_threads;
+    factx.n_threads      = n_threads;
     factx.DK             = DK;
     factx.DV             = DV;
     factx.n_kv           = nek1;
@@ -1328,10 +1325,15 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
     factx.m1          = powf(2.0f, -(max_bias / 2.0f) / factx.n_head_log2);
 
     // ======== VTCM allocation (GQA-aware) ========
+    const size_t size_k_row        = DK * sizeof(__fp16);
+    const size_t size_v_row        = DV * sizeof(__fp16);
+    const size_t size_k_row_padded = hex_round_up(size_k_row, 128);
+    const size_t size_v_row_padded = hex_round_up(size_v_row, 128);
+
     const size_t q_tile_bytes  = hex_align_up(g_br * DK * sizeof(__fp16), 4096);
     const size_t o_tile_bytes  = hex_align_up(g_br * DV * sizeof(__fp16), 4096);
-    const size_t k_dma_bytes   = hex_align_up(Bc * DK * sizeof(__fp16), 4096);
-    const size_t v_dma_bytes   = hex_align_up(Bc * DV * sizeof(__fp16), 4096);
+    const size_t k_dma_bytes   = hex_align_up(Bc * size_k_row_padded, 4096);
+    const size_t v_dma_bytes   = hex_align_up(Bc * size_v_row_padded, 4096);
     const size_t k_tile_bytes  = hex_align_up(Bc * DK * sizeof(__fp16), 4096);
     const size_t v_tile_bytes  = hex_align_up(Bc * DV * sizeof(__fp16), 4096);
     const size_t s_tile_bytes  = hex_align_up(g_br * Bc * sizeof(__fp16), 4096);
@@ -1401,11 +1403,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
     // ======== DMA setup ========
     dma_queue * const dma = ctx->dma[0];
 
-    // Padded row sizes for DMA
-    const size_t size_k_row        = nek0 * sizeof(__fp16);
-    const size_t size_v_row        = nev0 * sizeof(__fp16);
-    const size_t size_k_row_padded = hex_round_up(nek0 * sizeof(__fp16), 128);
-    const size_t size_v_row_padded = hex_round_up(nev0 * sizeof(__fp16), 128);
+    // Padded row sizes for DMA (defined in outer scope)
 
     const size_t n_row_tiles_g_br = g_br / HMX_FP16_TILE_N_ROWS;
     const size_t n_tiles_per_bc   = Bc / HMX_FP16_TILE_N_COLS;

ggml/src/ggml-hexagon/htp/hmx-matmul-ops.c

@@ -16,6 +16,7 @@
 #include "ggml-common.h"
 
 #include "hex-dma.h"
+#include "hex-fastdiv.h"
 #include "worker-pool.h"
 
 #include "hvx-utils.h"
@@ -34,6 +35,10 @@ static const __fp16 q4_0_to_fp16_lut[64] __attribute__((aligned(VLEN))) = {
     -8, 0, -7, 0, -6, 0, -5, 0, -4, 0, -3, 0, -2, 0, -1, 0, 0, 0, 1, 0, 2, 0, 3, 0, 4, 0, 5, 0, 6, 0, 7, 0,
 };
 
+static const __fp16 q4_1_to_fp16_lut[64] __attribute__((aligned(VLEN))) = {
+    0, 0, 1, 0, 2, 0, 3, 0, 4, 0, 5, 0, 6, 0, 7, 0, 8, 0, 9, 0, 10, 0, 11, 0, 12, 0, 13, 0, 14, 0, 15, 0,
+};
+
 // MXFP4 dequantization LUT: maps 4-bit index to fp16 mantissa value
 // kvalues: 0, 0.5, 1, 1.5, 2, 3, 4, 6, 0, -0.5, -1, -1.5, -2, -3, -4, -6
 static const __fp16 mxfp4_to_fp16_lut[64] __attribute__((aligned(VLEN))) = {
@@ -62,6 +67,8 @@ static inline size_t get_x4x2_row_stride(int weight_type, int k) {
         case HTP_TYPE_Q4_0:
         case HTP_TYPE_IQ4_NL:
             return (size_t) nb * (QK_Q4_0x4x2 / 2 + HMX_X4X2_DBLK_SIZE);         // 144 * nb
+        case HTP_TYPE_Q4_1:
+            return (size_t) nb * (QK_Q4_0x4x2 / 2 + 32);                         // 160 * nb
         case HTP_TYPE_Q8_0:
             return (size_t) nb * (QK_Q8_0x4x2 + HMX_X4X2_DBLK_SIZE);             // 272 * nb
         case HTP_TYPE_MXFP4:
@@ -181,45 +188,44 @@ next_nc:
 // In x4x2, sub-blocks 0..3 use lower nibbles, sub-blocks 4..7 use upper nibbles
 // of the same 32 packed bytes.
 static inline HVX_Vector dequantize_x4x2_q4_0_group_hvx(const uint8_t *packed_32, bool upper_nibbles, const __fp16 *scale, const HVX_Vector vlut_cvt) {
+    (void)vlut_cvt;
     HVX_Vector vq = hvx_vmemu(packed_32);
     const HVX_Vector mask_h4 = Q6_Vb_vsplat_R(0x0F);
+    const HVX_Vector i8 = Q6_Vb_vsplat_R(8);
     HVX_Vector v_scales = hvx_vec_repl_f16(hvx_vmemu(scale));
-    // q4x4x2 stores two int4 values per byte. Keep only the selected nibble.
-    HVX_Vector v_quants =  Q6_Vub_vlsr_VubR(vq, 4 * upper_nibbles);
+
+    HVX_Vector v_quants = Q6_Vub_vlsr_VubR(vq, 4 * upper_nibbles);
     v_quants = Q6_V_vand_VV(v_quants, mask_h4);
-    // Shuffle before LUT
-    v_quants = Q6_Vb_vshuff_Vb(v_quants);
-    // Use standard vlut16 (not _nomatch) to avoid stale-register NaN.
-    // _nomatch retains the previous destination-register value for colliding
-    // indices, but the C intrinsic doesn't model the implicit read so the
-    // compiler may allocate a register containing garbage/NaN.
-    HVX_VectorPair vp = Q6_Wh_vlut16_VbVhR(v_quants, vlut_cvt, 0);
-    HVX_Vector v_hf = Q6_V_lo_W(vp);
+
+    HVX_Vector v_int8 = Q6_Vb_vsub_VbVb(v_quants, i8);
+    HVX_Vector v0     = Q6_V_lo_W(Q6_Wh_vunpack_Vb(v_int8));
+    HVX_Vector v_hf   = Q6_Vhf_equals_Vh(v0);
 
     return Q6_Vhf_equals_Vqf16(Q6_Vqf16_vmpy_VhfVhf(v_hf, v_scales));
 }
 
 // Batch-dequantize 4 contiguous x4x2 Q4_0 groups (4x32 = 128 packed bytes) using
-// full HVX vector width.  One vmemu + one vlut16 replaces 4 separate calls.
+// full HVX vector width.
 // Output: vector_x2 each hold 32 FP16 values in the first 64 bytes.
 static inline HVX_Vector_x2 dequantize_x4x2_q4_0_x4groups_hvx(
             const uint8_t *packed_128, bool upper_nibbles,
             const __fp16 *scales_4, const HVX_Vector vlut_cvt) {
-    // Load all 128 packed bytes (4 contiguous 32-byte groups)
+    (void)vlut_cvt;
     HVX_Vector vq = hvx_vmemu(packed_128);
     const HVX_Vector mask_h4 = Q6_Vb_vsplat_R(0x0F);
+    const HVX_Vector i8 = Q6_Vb_vsplat_R(8);
     HVX_Vector v_quants = Q6_Vub_vlsr_VubR(vq, 4 * upper_nibbles);
     v_quants = Q6_V_vand_VV(v_quants, mask_h4);
 
-    // Shuffle before LUT
-    v_quants = Q6_Vb_vshuff_Vb(v_quants);
+    HVX_Vector v_int8 = Q6_Vb_vsub_VbVb(v_quants, i8);
 
-    // Full-width vlut16: 128 byte lookups -> 128 fp16 results in a VectorPair
-    HVX_VectorPair vp = Q6_Wh_vlut16_VbVhR(v_quants, vlut_cvt, 0);
-    HVX_Vector v_lo = Q6_V_lo_W(vp);  // [group0: 32 fp16 | group1: 32 fp16]
-    HVX_Vector v_hi = Q6_V_hi_W(vp);  // [group2: 32 fp16 | group3: 32 fp16]
+    HVX_VectorPair vp_int16 = Q6_Wh_vunpack_Vb(v_int8);
+    HVX_Vector v_lo = Q6_V_lo_W(vp_int16);
+    HVX_Vector v_hi = Q6_V_hi_W(vp_int16);
+
+    v_lo = Q6_Vhf_equals_Vh(v_lo);
+    v_hi = Q6_Vhf_equals_Vh(v_hi);
 
-    // Build per-group scale vectors: first 64 bytes use scale_a, last 64 use scale_b
     HVX_Vector vscale = hvx_vmemu(scales_4);
     HVX_Vector v_sc01 = hvx_vec_repl_2x_f16(vscale);
     HVX_Vector v_sc23 = hvx_vec_repl_2x_f16(Q6_V_vror_VR(vscale, 4));
@@ -227,9 +233,97 @@ static inline HVX_Vector_x2 dequantize_x4x2_q4_0_x4groups_hvx(
     v_lo = Q6_Vhf_equals_Vqf16(Q6_Vqf16_vmpy_VhfVhf(v_lo, v_sc01));
     v_hi = Q6_Vhf_equals_Vqf16(Q6_Vqf16_vmpy_VhfVhf(v_hi, v_sc23));
 
-    // Extract individual groups: scatter uses q_mask64 so only first 64 bytes matter
-    HVX_Vector_x2 r = { v_lo,/* group1 already in [0:63] */
-                        v_hi /* group2 already in [0:63] */ };
+    HVX_Vector_x2 r = { v_lo, v_hi };
+    return r;
+}
+
+static inline HVX_Vector dequantize_x4x2_q4_1_group_hvx(const uint8_t *packed_32, bool upper_nibbles, const __fp16 *scale_offset, const HVX_Vector vlut_cvt) {
+    (void)vlut_cvt;
+    HVX_Vector vq = hvx_vmemu(packed_32);
+    const HVX_Vector mask_h4 = Q6_Vb_vsplat_R(0x0F);
+    HVX_Vector v_dm = hvx_vmemu(scale_offset);
+    HVX_Vector v_scales = hvx_vec_repl_f16(v_dm);
+    HVX_Vector v_offsets = hvx_vec_repl_f16(Q6_V_vror_VR(v_dm, 2));
+
+    HVX_Vector v_quants =  Q6_Vub_vlsr_VubR(vq, 4 * upper_nibbles);
+    v_quants = Q6_V_vand_VV(v_quants, mask_h4);
+
+    HVX_Vector v0   = Q6_V_lo_W(Q6_Wh_vunpack_Vb(v_quants));
+    HVX_Vector v_hf = Q6_Vhf_equals_Vh(v0);
+
+    return Q6_Vhf_equals_Vqf16(Q6_Vqf16_vadd_Vqf16Vhf(Q6_Vqf16_vmpy_VhfVhf(v_hf, v_scales), v_offsets));
+}
+
+static inline HVX_Vector_x2 dequantize_x4x2_q4_1_x4groups_hvx(
+            const uint8_t *packed_128, bool upper_nibbles,
+            const __fp16 *scales_offsets_4, const HVX_Vector vlut_cvt) {
+    (void)vlut_cvt;
+    HVX_Vector vq = hvx_vmemu(packed_128);
+    const HVX_Vector mask_h4 = Q6_Vb_vsplat_R(0x0F);
+    HVX_Vector v_quants = Q6_Vub_vlsr_VubR(vq, 4 * upper_nibbles);
+    v_quants = Q6_V_vand_VV(v_quants, mask_h4);
+
+    HVX_VectorPair vp_int16 = Q6_Wh_vunpack_Vb(v_quants);
+    HVX_Vector v_lo = Q6_V_lo_W(vp_int16);
+    HVX_Vector v_hi = Q6_V_hi_W(vp_int16);
+
+    v_lo = Q6_Vhf_equals_Vh(v_lo);
+    v_hi = Q6_Vhf_equals_Vh(v_hi);
+
+    HVX_Vector vscale_offset = hvx_vmemu(scales_offsets_4);
+    HVX_VectorPair dm_deal = Q6_W_vdeal_VVR(vscale_offset, vscale_offset, -2);
+    HVX_Vector vd = Q6_V_lo_W(dm_deal);
+    HVX_Vector vm = Q6_V_hi_W(dm_deal);
+
+    HVX_Vector v_sc01 = hvx_vec_repl_2x_f16(vd);
+    HVX_Vector v_sc23 = hvx_vec_repl_2x_f16(Q6_V_vror_VR(vd, 4));
+
+    HVX_Vector v_os01 = hvx_vec_repl_2x_f16(vm);
+    HVX_Vector v_os23 = hvx_vec_repl_2x_f16(Q6_V_vror_VR(vm, 4));
+
+    v_lo = Q6_Vhf_equals_Vqf16(Q6_Vqf16_vadd_Vqf16Vhf(Q6_Vqf16_vmpy_VhfVhf(v_lo, v_sc01), v_os01));
+    v_hi = Q6_Vhf_equals_Vqf16(Q6_Vqf16_vadd_Vqf16Vhf(Q6_Vqf16_vmpy_VhfVhf(v_hi, v_sc23), v_os23));
+
+    HVX_Vector_x2 r = { v_lo, v_hi };
+    return r;
+}
+
+// LUT-based dequantizers for non-linear IQ4_NL format.
+static inline HVX_Vector dequantize_x4x2_iq4_nl_group_hvx(const uint8_t *packed_32, bool upper_nibbles, const __fp16 *scale, const HVX_Vector vlut_cvt) {
+    HVX_Vector vq = hvx_vmemu(packed_32);
+    const HVX_Vector mask_h4 = Q6_Vb_vsplat_R(0x0F);
+    HVX_Vector v_scales = hvx_vec_repl_f16(hvx_vmemu(scale));
+    HVX_Vector v_quants =  Q6_Vub_vlsr_VubR(vq, 4 * upper_nibbles);
+    v_quants = Q6_V_vand_VV(v_quants, mask_h4);
+    v_quants = Q6_Vb_vshuff_Vb(v_quants);
+    HVX_VectorPair vp = Q6_Wh_vlut16_VbVhR(v_quants, vlut_cvt, 0);
+    HVX_Vector v_hf = Q6_V_lo_W(vp);
+
+    return Q6_Vhf_equals_Vqf16(Q6_Vqf16_vmpy_VhfVhf(v_hf, v_scales));
+}
+
+static inline HVX_Vector_x2 dequantize_x4x2_iq4_nl_x4groups_hvx(
+            const uint8_t *packed_128, bool upper_nibbles,
+            const __fp16 *scales_4, const HVX_Vector vlut_cvt) {
+    HVX_Vector vq = hvx_vmemu(packed_128);
+    const HVX_Vector mask_h4 = Q6_Vb_vsplat_R(0x0F);
+    HVX_Vector v_quants = Q6_Vub_vlsr_VubR(vq, 4 * upper_nibbles);
+    v_quants = Q6_V_vand_VV(v_quants, mask_h4);
+
+    v_quants = Q6_Vb_vshuff_Vb(v_quants);
+
+    HVX_VectorPair vp = Q6_Wh_vlut16_VbVhR(v_quants, vlut_cvt, 0);
+    HVX_Vector v_lo = Q6_V_lo_W(vp);
+    HVX_Vector v_hi = Q6_V_hi_W(vp);
+
+    HVX_Vector vscale = hvx_vmemu(scales_4);
+    HVX_Vector v_sc01 = hvx_vec_repl_2x_f16(vscale);
+    HVX_Vector v_sc23 = hvx_vec_repl_2x_f16(Q6_V_vror_VR(vscale, 4));
+
+    v_lo = Q6_Vhf_equals_Vqf16(Q6_Vqf16_vmpy_VhfVhf(v_lo, v_sc01));
+    v_hi = Q6_Vhf_equals_Vqf16(Q6_Vqf16_vmpy_VhfVhf(v_hi, v_sc23));
+
+    HVX_Vector_x2 r = { v_lo, v_hi };
     return r;
 }
 
@@ -320,100 +414,176 @@ static inline HVX_Vector_x4 dequantize_x4x2_mxfp4_x4groups_hvx(const uint8_t *
     return r;
 }
 
+typedef struct {
+    __fp16                  *dst;
+    const uint8_t           *src;
+    int                      n_cols;
+    int                      k_block;
+    size_t                   row_stride;
+    int                      weight_type;
+    int                      n_tot_tiles;
+    int                      n_tiles_per_task;
+    int                      n_tasks;
+    int                      n_k_tiles;
+    struct fastdiv_values    n_k_tiles_div;
+} x4x2_dequantize_state_t;
+
 // Dequantize a tile range from x4x2 weight data (already in VTCM) to tile-major FP16.
 // Input:  vtcm_src has n_cols rows of x4x2 data, each row_stride bytes.
 // Output: vtcm_dst in tile-major FP16 layout.
-static void dequantize_x4x2_weight_to_fp16_tiles_task(
-        __fp16 *restrict vtcm_dst,
-        const uint8_t *restrict vtcm_src,
-        int n_cols, int k_block,
-        size_t row_stride, int weight_type,
+
+#define DEFINE_DEQUANTIZE_Q4_TASK(suffix, lut_name, helper_prefix, dblk_size, scale_step)                      \
+static void dequantize_x4x2_weight_to_fp16_tiles_task_##suffix(                                                \
+        const x4x2_dequantize_state_t *state,                                                                  \
+        int start_tile, int end_tile) {                                                                        \
+                                                                                                               \
+    const int n_k_tiles = state->n_k_tiles;                                                                    \
+    const int qrow_size = (unsigned)state->k_block / 2;                                                        \
+    const struct fastdiv_values n_k_tiles_div = state->n_k_tiles_div;                                          \
+    const HVX_Vector vlut_cvt = hvx_vmem(lut_name);                                                            \
+                                                                                                               \
+    const HVX_Vector v_scat_base  = hvx_vmem(hmx_transpose_scatter_offsets);                                   \
+    const HVX_Vector v_scat_step  = Q6_V_vsplat_R(4);                                                          \
+    const HVX_VectorPred q_mask64 = Q6_Q_vsetq_R(64);                                                          \
+                                                                                                               \
+    unsigned ct = fastdiv((unsigned)start_tile, &n_k_tiles_div);                                               \
+    unsigned kt = fastmodulo((unsigned)start_tile, n_k_tiles, &n_k_tiles_div);                                 \
+                                                                                                               \
+    for (unsigned t = start_tile; t < (unsigned)end_tile; ) {                                                  \
+        if (kt >= (unsigned)n_k_tiles) { kt = 0; ct++; }                                                       \
+                                                                                                               \
+        if ((kt % 4 == 0) && (t + 4 <= (unsigned)end_tile) && (fastdiv(t + 3, &n_k_tiles_div) == ct)) {        \
+            unsigned blk_idx      = ((kt * 32) / QK_Q4_0x4x2);                                                 \
+            unsigned sub_blk_base = ((kt * 32) % QK_Q4_0x4x2) / 32;                                            \
+            bool upper            = (sub_blk_base >= 4);                                                       \
+            unsigned packed_off   = blk_idx * (QK_Q4_0x4x2 / 2);                                               \
+            unsigned scale_off    = qrow_size + blk_idx * (dblk_size) + sub_blk_base * (scale_step);           \
+                                                                                                               \
+            __fp16 *tile_bases[4];                                                                             \
+            for (unsigned g = 0; g < 4; g++) {                                                                 \
+                tile_bases[g] = state->dst + (t + g) * HMX_FP16_TILE_N_ELMS;                                   \
+            }                                                                                                  \
+                                                                                                               \
+            HVX_Vector v_off = v_scat_base;                                                                    \
+            unsigned row_offset = ct * HMX_FP16_TILE_N_COLS * state->row_stride;                               \
+                                                                                                               \
+            for (int r = 0; r < HMX_FP16_TILE_N_ROWS; r += 2) {                                                \
+                const uint8_t *r0 = state->src + row_offset; row_offset += state->row_stride;                  \
+                const uint8_t *r1 = state->src + row_offset; row_offset += state->row_stride;                  \
+                                                                                                               \
+                HVX_Vector_x2 dv0 = dequantize_x4x2_##helper_prefix##_x4groups_hvx(                            \
+                    r0 + packed_off, upper, (const __fp16 *)(r0 + scale_off), vlut_cvt);                       \
+                Q6_vscatter_RMVwV((size_t)tile_bases[0], 2 * HMX_FP16_TILE_SIZE - 1, v_off, dv0.v[0]);         \
+                Q6_vscatter_RMVwV((size_t)tile_bases[2], 2 * HMX_FP16_TILE_SIZE - 1, v_off, dv0.v[1]);         \
+                v_off = Q6_Vw_vadd_VwVw(v_off, v_scat_step);                                                   \
+                                                                                                               \
+                HVX_Vector_x2 dv1 = dequantize_x4x2_##helper_prefix##_x4groups_hvx(                            \
+                    r1 + packed_off, upper, (const __fp16 *)(r1 + scale_off), vlut_cvt);                       \
+                Q6_vscatter_RMVwV((size_t)tile_bases[0], 2 * HMX_FP16_TILE_SIZE - 1, v_off, dv1.v[0]);         \
+                Q6_vscatter_RMVwV((size_t)tile_bases[2], 2 * HMX_FP16_TILE_SIZE - 1, v_off, dv1.v[1]);         \
+                v_off = Q6_Vw_vadd_VwVw(v_off, v_scat_step);                                                   \
+            }                                                                                                  \
+                                                                                                               \
+            for (int g = 0; g < 4; g++) { (void) *(volatile HVX_Vector *)(tile_bases[g]); }                    \
+            t += 4; kt += 4;                                                                                   \
+            continue;                                                                                          \
+        }                                                                                                      \
+                                                                                                               \
+        __fp16 *tile_base = state->dst + t * HMX_FP16_TILE_N_ELMS;                                             \
+        {                                                                                                      \
+            unsigned blk_idx   = (kt * 32) / QK_Q4_0x4x2;                                                      \
+            unsigned sub_blk   = ((kt * 32) % QK_Q4_0x4x2) / 32;                                               \
+            bool upper         = (sub_blk >= 4);                                                               \
+            unsigned byte_off  = blk_idx * (QK_Q4_0x4x2 / 2) + (upper ? (sub_blk - 4) : sub_blk) * 32;         \
+            unsigned scale_off = qrow_size + blk_idx * (dblk_size) + sub_blk * (scale_step);                   \
+                                                                                                               \
+            HVX_Vector v_off = v_scat_base;                                                                    \
+            unsigned row_offset = ct * HMX_FP16_TILE_N_COLS * state->row_stride;                               \
+            unsigned row1 = ct * HMX_FP16_TILE_N_COLS + 1;                                                     \
+                                                                                                               \
+            for (int r = 0; r < HMX_FP16_TILE_N_ROWS; r += 2, row1 += 2) {                                     \
+                const uint8_t *r0 = state->src + row_offset; row_offset += state->row_stride;                  \
+                const uint8_t *r1 = state->src + row_offset; row_offset += state->row_stride;                  \
+                                                                                                               \
+                HVX_Vector v0 = dequantize_x4x2_##helper_prefix##_group_hvx(                                   \
+                    r0 + byte_off, upper, (const __fp16 *)(r0 + scale_off), vlut_cvt);                         \
+                HVX_Vector v1 = (row1 < (unsigned)state->n_cols)                                               \
+                    ? dequantize_x4x2_##helper_prefix##_group_hvx(                                             \
+                        r1 + byte_off, upper, (const __fp16 *)(r1 + scale_off), vlut_cvt)                      \
+                    : Q6_V_vzero();                                                                            \
+                                                                                                               \
+                Q6_vscatter_QRMVwV(q_mask64, (size_t)tile_base, HMX_FP16_TILE_SIZE - 1, v_off, v0);            \
+                v_off = Q6_Vw_vadd_VwVw(v_off, v_scat_step);                                                   \
+                Q6_vscatter_QRMVwV(q_mask64, (size_t)tile_base, HMX_FP16_TILE_SIZE - 1, v_off, v1);            \
+                v_off = Q6_Vw_vadd_VwVw(v_off, v_scat_step);                                                   \
+            }                                                                                                  \
+            (void) *(volatile HVX_Vector *)(tile_base);                                                        \
+        }                                                                                                      \
+        ++t; ++kt;                                                                                             \
+    }                                                                                                          \
+                                                                                                               \
+    if (start_tile < end_tile) {                                                                               \
+        (void) *(volatile HVX_Vector *)(state->dst + (end_tile - 1) * HMX_FP16_TILE_N_ELMS);                   \
+    }                                                                                                          \
+}                                                                                                              \
+                                                                                                               \
+static void dequantize_x4x2_worker_loop_##suffix(unsigned int n, unsigned int i, void *data) {                 \
+    x4x2_dequantize_state_t *state = (x4x2_dequantize_state_t *)data;                                          \
+    for (unsigned int task_id = i; task_id < (unsigned int)state->n_tasks; task_id += n) {                     \
+        int start = task_id * state->n_tiles_per_task;                                                         \
+        int end   = hex_smin(start + state->n_tiles_per_task, state->n_tot_tiles);                             \
+        dequantize_x4x2_weight_to_fp16_tiles_task_##suffix(state, start, end);                                 \
+    }                                                                                                          \
+}
+
+DEFINE_DEQUANTIZE_Q4_TASK(q4_0,   q4_0_to_fp16_lut,   q4_0, HMX_X4X2_DBLK_SIZE, (int)sizeof(__fp16))
+DEFINE_DEQUANTIZE_Q4_TASK(q4_1,   q4_1_to_fp16_lut,   q4_1, 32, 4)
+DEFINE_DEQUANTIZE_Q4_TASK(iq4_nl, iq4_nl_to_fp16_lut, iq4_nl, HMX_X4X2_DBLK_SIZE, (int)sizeof(__fp16))
+
+static void dequantize_x4x2_weight_to_fp16_tiles_task_mxfp4(
+        const x4x2_dequantize_state_t *state,
         int start_tile, int end_tile) {
 
-    const int n_k_tiles = (unsigned)k_block / HMX_FP16_TILE_N_COLS;
-    const bool is_q4 = (weight_type == HTP_TYPE_Q4_0 || weight_type == HTP_TYPE_IQ4_NL);
-    const int qrow_size = is_q4 ? ((unsigned)k_block / 2) : k_block;
+    const int n_k_tiles = state->n_k_tiles;
+    const int qrow_size = state->k_block;
+    const struct fastdiv_values n_k_tiles_div = state->n_k_tiles_div;
+    const HVX_Vector vlut_cvt = hvx_vmem(mxfp4_to_fp16_lut);
 
-    const HVX_Vector vlut_cvt = (weight_type == HTP_TYPE_IQ4_NL) ? hvx_vmem(iq4_nl_to_fp16_lut) :
-                                (weight_type == HTP_TYPE_MXFP4)  ? hvx_vmem(mxfp4_to_fp16_lut) :
-                                                                   hvx_vmem(q4_0_to_fp16_lut);
-
-    // vscatter setup: write dequantized K-values directly to transposed [K][N] tile positions.
-    // Each int32 element holds a K-row-pair (2 adjacent fp16 values).  word[i] at offset i*128
-    // maps to K-rows 2i and 2i+1.  Column offset (n*4) added per row.
     const HVX_Vector v_scat_base  = hvx_vmem(hmx_transpose_scatter_offsets);
-    const HVX_Vector v_scat_step  = Q6_V_vsplat_R(4);  // 4 bytes = 1 column step
-    const HVX_VectorPred q_mask64 = Q6_Q_vsetq_R(64);  // first 16 words (64 bytes)
+    const HVX_Vector v_scat_step  = Q6_V_vsplat_R(4);
+    const HVX_VectorPred q_mask64 = Q6_Q_vsetq_R(64);
 
-    unsigned ct = (unsigned)start_tile / n_k_tiles;  // column tile index
-    unsigned kt = (unsigned)start_tile % n_k_tiles;  // K tile index
-    for (unsigned t = start_tile; t < end_tile; ) {
-        if (kt >= n_k_tiles) { kt = 0; ct++; }
+    unsigned ct = fastdiv((unsigned)start_tile, &n_k_tiles_div);
+    unsigned kt = fastmodulo((unsigned)start_tile, n_k_tiles, &n_k_tiles_div);
 
-        // --- Batch-4 fast path for Q4: process 4 contiguous K-tiles with one vlut16 per row ---
-        if (is_q4 && (kt % 4 == 0) && (t + 4 <= end_tile) && ((t + 3) / n_k_tiles == ct)) {
-            unsigned blk_idx      = (kt * 32) / QK_Q4_0x4x2;
-            unsigned sub_blk_base = ((kt * 32) % QK_Q4_0x4x2) / 32;  // 0 or 4
-            bool upper            = (sub_blk_base >= 4);
-            unsigned packed_off   = blk_idx * (QK_Q4_0x4x2 / 2);     // 128 contiguous packed bytes
-            unsigned scale_off    = qrow_size + blk_idx * HMX_X4X2_DBLK_SIZE
-                                  + sub_blk_base * (int)sizeof(__fp16);   // 4 consecutive scales
+    for (unsigned t = start_tile; t < (unsigned)end_tile; ) {
+        if (kt >= (unsigned)n_k_tiles) { kt = 0; ct++; }
 
-            __fp16 *tile_bases[4];
-            for (unsigned g = 0; g < 4; g++) { tile_bases[g] = vtcm_dst + (t + g) * HMX_FP16_TILE_N_ELMS; }
-
-            HVX_Vector v_off = v_scat_base;
-
-            unsigned row_offset = ct * HMX_FP16_TILE_N_COLS * row_stride;
-            unsigned row1 = ct * HMX_FP16_TILE_N_COLS + 1;
-
-            for (int r = 0; r < HMX_FP16_TILE_N_ROWS; r += 2, row1 += 2) {
-                const uint8_t *r0 = vtcm_src + row_offset; row_offset += row_stride;
-                const uint8_t *r1 = vtcm_src + row_offset; row_offset += row_stride;
-
-                HVX_Vector_x2 dv0 = dequantize_x4x2_q4_0_x4groups_hvx(r0 + packed_off, upper, (const __fp16 *)(r0 + scale_off), vlut_cvt);
-                HVX_Vector_x2 dv1 = dequantize_x4x2_q4_0_x4groups_hvx(r1 + packed_off, upper, (const __fp16 *)(r1 + scale_off), vlut_cvt);
-
-                Q6_vscatter_RMVwV((size_t)tile_bases[0], 2 * HMX_FP16_TILE_SIZE - 1, v_off, dv0.v[0]);
-                Q6_vscatter_RMVwV((size_t)tile_bases[2], 2 * HMX_FP16_TILE_SIZE - 1, v_off, dv0.v[1]);
-                v_off = Q6_Vw_vadd_VwVw(v_off, v_scat_step);
-
-                Q6_vscatter_RMVwV((size_t)tile_bases[0], 2 * HMX_FP16_TILE_SIZE - 1, v_off, dv1.v[0]);
-                Q6_vscatter_RMVwV((size_t)tile_bases[2], 2 * HMX_FP16_TILE_SIZE - 1, v_off, dv1.v[1]);
-                v_off = Q6_Vw_vadd_VwVw(v_off, v_scat_step);
-            }
-
-            for (int g = 0; g < 4; g++) { (void) *(volatile HVX_Vector *)(tile_bases[g]); }
-            t += 4; kt += 4;
-            continue;
-        }
-
-        // --- Batch-4 fast path for MXFP4: same nibble layout but E8M0 scales ---
-        if (weight_type == HTP_TYPE_MXFP4 && (kt % 4 == 0) && (t + 4 <= end_tile) && ((t + 3) / n_k_tiles == ct)) {
+        // Batch-4 fast path for MXFP4
+        if ((kt % 4 == 0) && (t + 4 <= (unsigned)end_tile) && (fastdiv(t + 3, &n_k_tiles_div) == ct)) {
             int  blk_idx      = (kt * 32) / QK_MXFP4x4x2;
-            int  sub_blk_base = ((kt * 32) % QK_MXFP4x4x2) / 32;                 // 0 or 4
+            int  sub_blk_base = ((kt * 32) % QK_MXFP4x4x2) / 32;
             bool upper        = (sub_blk_base >= 4);
-            int  packed_off   = blk_idx * (QK_MXFP4x4x2 / 2);                    // 128 contiguous packed bytes
-            int  e8m0_blk_off = qrow_size + blk_idx * HMX_X4X2_MXFP4_EBLK_SIZE;  // all 8 E8M0 scales
+            int  packed_off   = blk_idx * (QK_MXFP4x4x2 / 2);
+            int  e8m0_blk_off = qrow_size + blk_idx * HMX_X4X2_MXFP4_EBLK_SIZE;
 
             __fp16 * tile_bases[4];
             for (int g = 0; g < 4; g++) {
-                tile_bases[g] = vtcm_dst + (t + g) * HMX_FP16_TILE_N_ELMS;
+                tile_bases[g] = state->dst + (t + g) * HMX_FP16_TILE_N_ELMS;
             }
 
             HVX_Vector v_off = v_scat_base;
             for (int r = 0; r < HMX_FP16_TILE_N_ROWS; r += 2) {
                 int             row0 = ct * HMX_FP16_TILE_N_COLS + r;
                 int             row1 = row0 + 1;
-                const uint8_t * r0   = vtcm_src + row0 * row_stride;
-                const uint8_t * r1   = vtcm_src + row1 * row_stride;
+                const uint8_t * r0   = state->src + row0 * state->row_stride;
+                const uint8_t * r1   = state->src + row1 * state->row_stride;
 
-                // Batch-convert all 8 E8M0 scales once per row (stays in HVX register)
                 mxfp4_scales_t r0_e8 = mxfp4_convert_scales(r0 + e8m0_blk_off);
 
                 HVX_Vector_x4 dv0, dv1;
                 dv0 = dequantize_x4x2_mxfp4_x4groups_hvx(r0 + packed_off, upper, sub_blk_base, vlut_cvt, r0_e8);
-                if (row1 < n_cols) {
+                if (row1 < state->n_cols) {
                     mxfp4_scales_t r1_e8 = mxfp4_convert_scales(r1 + e8m0_blk_off);
                     dv1 = dequantize_x4x2_mxfp4_x4groups_hvx(r1 + packed_off, upper, sub_blk_base, vlut_cvt, r1_e8);
                 } else {
@@ -434,41 +604,13 @@ static void dequantize_x4x2_weight_to_fp16_tiles_task(
                 (void) *(volatile HVX_Vector *) (tile_bases[g]);
             }
 
-            t += 4;
+            t += 4; kt += 4;
             continue;
         }
 
-        // --- Single-tile fallback ---
-        __fp16 *tile_base = vtcm_dst + t * HMX_FP16_TILE_N_ELMS;
-
-        if (is_q4) {
-            unsigned blk_idx   = (kt * 32) / QK_Q4_0x4x2;
-            unsigned sub_blk   = ((kt * 32) % QK_Q4_0x4x2) / 32;
-            bool upper         = (sub_blk >= 4);
-            unsigned byte_off  = blk_idx * (QK_Q4_0x4x2 / 2) + (upper ? (sub_blk - 4) : sub_blk) * 32;
-            unsigned scale_off = qrow_size + blk_idx * HMX_X4X2_DBLK_SIZE + sub_blk * (int)sizeof(__fp16);
-
-            HVX_Vector v_off = v_scat_base;  // reset to column 0
-            unsigned row_offset = ct * HMX_FP16_TILE_N_COLS * row_stride;
-            unsigned row1 = ct * HMX_FP16_TILE_N_COLS + 1;
-            for (int r = 0; r < HMX_FP16_TILE_N_ROWS; r += 2, row1 += 2) {
-                const uint8_t *r0 = vtcm_src + row_offset; row_offset += row_stride;
-                const uint8_t *r1 = vtcm_src + row_offset; row_offset += row_stride;
-
-                HVX_Vector v0 = dequantize_x4x2_q4_0_group_hvx(
-                    r0 + byte_off, upper, (const __fp16 *)(r0 + scale_off), vlut_cvt);
-                HVX_Vector v1 = (row1 < n_cols)
-                    ? dequantize_x4x2_q4_0_group_hvx(
-                        r1 + byte_off, upper, (const __fp16 *)(r1 + scale_off), vlut_cvt)
-                    : Q6_V_vzero();
-
-                Q6_vscatter_QRMVwV(q_mask64, (size_t)tile_base, HMX_FP16_TILE_SIZE - 1, v_off, v0);
-                v_off = Q6_Vw_vadd_VwVw(v_off, v_scat_step);
-                Q6_vscatter_QRMVwV(q_mask64, (size_t)tile_base, HMX_FP16_TILE_SIZE - 1, v_off, v1);
-                v_off = Q6_Vw_vadd_VwVw(v_off, v_scat_step);
-            }
-            (void) *(volatile HVX_Vector *)(tile_base);
-        } else if (weight_type == HTP_TYPE_MXFP4) {
+        // Single-tile fallback
+        __fp16 *tile_base = state->dst + t * HMX_FP16_TILE_N_ELMS;
+        {
             int  blk_idx      = (kt * 32) / QK_MXFP4x4x2;
             int  sub_blk      = ((kt * 32) % QK_MXFP4x4x2) / 32;
             bool upper        = (sub_blk >= 4);
@@ -480,15 +622,14 @@ static void dequantize_x4x2_weight_to_fp16_tiles_task(
                 int row0 = ct * HMX_FP16_TILE_N_COLS + r;
                 int row1 = row0 + 1;
 
-                const uint8_t * r0 = vtcm_src + row0 * row_stride;
-                const uint8_t * r1 = vtcm_src + row1 * row_stride;
+                const uint8_t * r0 = state->src + row0 * state->row_stride;
+                const uint8_t * r1 = state->src + row1 * state->row_stride;
 
-                // Batch-convert all 8 E8M0 scales once per row (stays in HVX register)
                 mxfp4_scales_t r0_e8 = mxfp4_convert_scales(r0 + e8m0_blk_off);
 
                 HVX_Vector v0 = dequantize_x4x2_mxfp4_group_hvx(r0 + byte_off, upper, sub_blk, vlut_cvt, r0_e8);
                 HVX_Vector v1;
-                if (row1 < n_cols) {
+                if (row1 < state->n_cols) {
                     mxfp4_scales_t r1_e8 = mxfp4_convert_scales(r1 + e8m0_blk_off);
                     v1 = dequantize_x4x2_mxfp4_group_hvx(r1 + byte_off, upper, sub_blk, vlut_cvt, r1_e8);
                 } else {
@@ -501,23 +642,59 @@ static void dequantize_x4x2_weight_to_fp16_tiles_task(
                 v_off = Q6_Vw_vadd_VwVw(v_off, v_scat_step);
             }
             (void) *(volatile HVX_Vector *) (tile_base);
-        } else {
-            // Q8_0
+        }
+        ++t; ++kt;
+    }
+
+    if (start_tile < end_tile) {
+        (void) *(volatile HVX_Vector *)(state->dst + (end_tile - 1) * HMX_FP16_TILE_N_ELMS);
+    }
+}
+
+static void dequantize_x4x2_worker_loop_mxfp4(unsigned int n, unsigned int i, void *data) {
+    x4x2_dequantize_state_t *state = (x4x2_dequantize_state_t *)data;
+    for (unsigned int task_id = i; task_id < (unsigned int)state->n_tasks; task_id += n) {
+        int start = task_id * state->n_tiles_per_task;
+        int end   = hex_smin(start + state->n_tiles_per_task, state->n_tot_tiles);
+        dequantize_x4x2_weight_to_fp16_tiles_task_mxfp4(state, start, end);
+    }
+}
+
+static void dequantize_x4x2_weight_to_fp16_tiles_task_q8_0(
+        const x4x2_dequantize_state_t *state,
+        int start_tile, int end_tile) {
+
+    const int n_k_tiles = state->n_k_tiles;
+    const int qrow_size = state->k_block;
+    const struct fastdiv_values n_k_tiles_div = state->n_k_tiles_div;
+
+    const HVX_Vector v_scat_base  = hvx_vmem(hmx_transpose_scatter_offsets);
+    const HVX_Vector v_scat_step  = Q6_V_vsplat_R(4);
+    const HVX_VectorPred q_mask64 = Q6_Q_vsetq_R(64);
+
+    unsigned ct = fastdiv((unsigned)start_tile, &n_k_tiles_div);
+    unsigned kt = fastmodulo((unsigned)start_tile, n_k_tiles, &n_k_tiles_div);
+
+    for (unsigned t = start_tile; t < (unsigned)end_tile; ) {
+        if (kt >= (unsigned)n_k_tiles) { kt = 0; ct++; }
+
+        __fp16 *tile_base = state->dst + t * HMX_FP16_TILE_N_ELMS;
+        {
             int blk_idx  = (kt * 32) / QK_Q8_0x4x2;
             int sub_blk  = ((kt * 32) % QK_Q8_0x4x2) / 32;
             int byte_off  = blk_idx * QK_Q8_0x4x2 + sub_blk * 32;
             int scale_off = qrow_size + blk_idx * HMX_X4X2_DBLK_SIZE + sub_blk * (int)sizeof(__fp16);
 
-            HVX_Vector v_off = v_scat_base;  // reset to column 0
+            HVX_Vector v_off = v_scat_base;
             for (int r = 0; r < HMX_FP16_TILE_N_ROWS; r += 2) {
                 int row0 = ct * HMX_FP16_TILE_N_COLS + r;
                 int row1 = row0 + 1;
 
-                const uint8_t *r0 = vtcm_src + row0 * row_stride;
-                const uint8_t *r1 = vtcm_src + row1 * row_stride;
+                const uint8_t *r0 = state->src + row0 * state->row_stride;
+                const uint8_t *r1 = state->src + row1 * state->row_stride;
 
                 HVX_Vector v0 = dequantize_x4x2_q8_0_group_hvx((const int8_t *)(r0 + byte_off), (const __fp16 *)(r0 + scale_off));
-                HVX_Vector v1 = (row1 < n_cols) ? dequantize_x4x2_q8_0_group_hvx((const int8_t *)(r1 + byte_off), (const __fp16 *)(r1 + scale_off)) : Q6_V_vzero();
+                HVX_Vector v1 = (row1 < state->n_cols) ? dequantize_x4x2_q8_0_group_hvx((const int8_t *)(r1 + byte_off), (const __fp16 *)(r1 + scale_off)) : Q6_V_vzero();
 
                 Q6_vscatter_QRMVwV(q_mask64, (size_t)tile_base, HMX_FP16_TILE_SIZE - 1, v_off, v0);
                 v_off = Q6_Vw_vadd_VwVw(v_off, v_scat_step);
@@ -529,50 +706,31 @@ static void dequantize_x4x2_weight_to_fp16_tiles_task(
         ++t; ++kt;
     }
 
-    // Drain HVX scatter write buffer: a vmem load on the same HW thread retires
-    // all pending scatter entries to VTCM.  Without this, the main thread's HMX
-    // reads may see stale data because atomic_fetch_sub (release) only orders
-    // regular stores, not the HVX scatter buffer.
     if (start_tile < end_tile) {
-        (void) *(volatile HVX_Vector *)(vtcm_dst + (end_tile - 1) * HMX_FP16_TILE_N_ELMS);
+        (void) *(volatile HVX_Vector *)(state->dst + (end_tile - 1) * HMX_FP16_TILE_N_ELMS);
     }
 }
 
-typedef struct {
-    __fp16        *dst;
-    const uint8_t *src;
-    int            n_cols;
-    int            k_block;
-    size_t         row_stride;
-    int            weight_type;
-    int            n_tot_tiles;
-    int            n_tiles_per_task;
-    int            n_tasks;
-} x4x2_dequantize_state_t;
-
-static void dequantize_x4x2_worker_loop(unsigned int n, unsigned int i, void *data) {
+static void dequantize_x4x2_worker_loop_q8_0(unsigned int n, unsigned int i, void *data) {
     x4x2_dequantize_state_t *state = (x4x2_dequantize_state_t *)data;
-
     for (unsigned int task_id = i; task_id < (unsigned int)state->n_tasks; task_id += n) {
         int start = task_id * state->n_tiles_per_task;
         int end   = hex_smin(start + state->n_tiles_per_task, state->n_tot_tiles);
-
-        dequantize_x4x2_weight_to_fp16_tiles_task(
-            state->dst, state->src, state->n_cols, state->k_block,
-            state->row_stride, state->weight_type, start, end);
+        dequantize_x4x2_weight_to_fp16_tiles_task_q8_0(state, start, end);
     }
 }
 
 static void dequantize_x4x2_weight_chunk_to_fp16_tiles(
         struct htp_context *ctx, __fp16 *vtcm_dst,
         const void *vtcm_src, int n_cols, int k_block,
-        size_t row_stride, int weight_type) {
+        size_t row_stride, int weight_type,
+        int n_k_tiles, struct fastdiv_values n_k_tiles_div,
+        worker_callback_t dequant_worker_fn) {
 
     assert(n_cols  % HMX_FP16_TILE_N_COLS == 0);
     assert(k_block % HMX_FP16_TILE_N_COLS == 0);
 
     size_t n_col_tiles = n_cols / HMX_FP16_TILE_N_COLS;
-    size_t n_k_tiles   = k_block / HMX_FP16_TILE_N_COLS;
     size_t n_tot_tiles = n_col_tiles * n_k_tiles;
 
     size_t n_tiles_per_task = hmx_ceil_div(n_tot_tiles, ctx->n_threads);
@@ -587,12 +745,16 @@ static void dequantize_x4x2_weight_chunk_to_fp16_tiles(
     state.k_block          = k_block;
     state.row_stride       = row_stride;
     state.weight_type      = weight_type;
+    state.n_k_tiles        = n_k_tiles;
+    state.n_k_tiles_div    = n_k_tiles_div;
 
-    worker_pool_run_func(ctx->worker_pool, dequantize_x4x2_worker_loop, &state, ctx->n_threads);
+    worker_pool_run_func(ctx->worker_pool, dequant_worker_fn, &state, ctx->n_threads);
 }
 
 // --- End x4x2 dequantizers ---
 
+#pragma clang diagnostic ignored "-Wbackend-plugin" // spurios warning for hmx intrinsics
+
 // requires external HMX lock
 static void core_dot_chunk_fp16(__fp16 *restrict output, const __fp16 *restrict activation, const __fp16 *restrict weight, const __fp16 *restrict scales,
                                 int n_row_tiles, int n_col_tiles, int n_dot_tiles) {
@@ -883,6 +1045,20 @@ int hmx_matmul_q_f32(struct htp_context *ctx, float *restrict dst, const float *
         return -1;
     }
 
+    worker_callback_t dequant_worker_fn = NULL;
+    switch (weight_type) {
+        case HTP_TYPE_Q4_0:   dequant_worker_fn = dequantize_x4x2_worker_loop_q4_0; break;
+        case HTP_TYPE_IQ4_NL: dequant_worker_fn = dequantize_x4x2_worker_loop_iq4_nl; break;
+        case HTP_TYPE_Q4_1:   dequant_worker_fn = dequantize_x4x2_worker_loop_q4_1; break;
+        case HTP_TYPE_MXFP4:  dequant_worker_fn = dequantize_x4x2_worker_loop_mxfp4; break;
+        case HTP_TYPE_Q8_0:   dequant_worker_fn = dequantize_x4x2_worker_loop_q8_0; break;
+        default:
+            return -1;
+    }
+
+    const int n_k_tiles = k / HMX_FP16_TILE_N_COLS;
+    const struct fastdiv_values n_k_tiles_div = init_fastdiv_values(n_k_tiles);
+
     // --- Dynamic VTCM layout ---
     const size_t vec_dot_size = k * sizeof(__fp16);
     const size_t vtcm_budget  = ctx->vtcm_size;
@@ -975,7 +1151,7 @@ int hmx_matmul_q_f32(struct htp_context *ctx, float *restrict dst, const float *
         {
             // B0: wait for DMA, dequant weight chunk 0
             dma_queue_pop(ctx->dma[0]);
-            dequantize_x4x2_weight_chunk_to_fp16_tiles(ctx, vtcm_weight_bufs[0], vtcm_qweight, n_cols_A0, k, row_stride, weight_type);
+            dequantize_x4x2_weight_chunk_to_fp16_tiles(ctx, vtcm_weight_bufs[0], vtcm_qweight, n_cols_A0, k, row_stride, weight_type, n_k_tiles, n_k_tiles_div, dequant_worker_fn);
 
             // A1: issue DMA for weight chunk 1
             const size_t n_cols_A1 = hex_smin(n - 1 * n_chunk_n_cols, n_chunk_n_cols);
@@ -994,7 +1170,7 @@ int hmx_matmul_q_f32(struct htp_context *ctx, float *restrict dst, const float *
             // B1: DMA pop + dequant (runs in parallel with C0 on HMX worker)
             if (1 < n_chunk_cnt) {
                 dma_queue_pop(ctx->dma[0]);
-                dequantize_x4x2_weight_chunk_to_fp16_tiles(ctx, vtcm_weight_bufs[1], vtcm_qweight, n_cols_A1, k, row_stride, weight_type);
+                dequantize_x4x2_weight_chunk_to_fp16_tiles(ctx, vtcm_weight_bufs[1], vtcm_qweight, n_cols_A1, k, row_stride, weight_type, n_k_tiles, n_k_tiles_div, dequant_worker_fn);
             }
         }
 
@@ -1036,7 +1212,7 @@ int hmx_matmul_q_f32(struct htp_context *ctx, float *restrict dst, const float *
             // B_{i+2}: DMA pop + dequant (multi-thread HVX, parallel with C_{i+1})
             if (i + 2 < n_chunk_cnt) {
                 dma_queue_pop(ctx->dma[0]);
-                dequantize_x4x2_weight_chunk_to_fp16_tiles(ctx, vtcm_weight_bufs[(i + 2) % 2], vtcm_qweight, n_cols_p2, k, row_stride, weight_type);
+                dequantize_x4x2_weight_chunk_to_fp16_tiles(ctx, vtcm_weight_bufs[(i + 2) % 2], vtcm_qweight, n_cols_p2, k, row_stride, weight_type, n_k_tiles, n_k_tiles_div, dequant_worker_fn);
             }
         }
     }

ggml/src/ggml-hexagon/htp/htp-ctx.h

@@ -104,6 +104,7 @@ int op_argsort(struct htp_ops_context * octx);
 int op_ssm_conv(struct htp_ops_context * octx);
 int op_cumsum(struct htp_ops_context * octx);
 int op_fill(struct htp_ops_context * octx);
+int op_concat(struct htp_ops_context * octx);
 int op_diag(struct htp_ops_context * octx);
 int op_solve_tri(struct htp_ops_context * octx);
 int op_gated_delta_net(struct htp_ops_context * octx);

ggml/src/ggml-hexagon/htp/htp-ops.h

@@ -20,6 +20,7 @@ enum htp_data_type {
     HTP_TYPE_F32    = 0,
     HTP_TYPE_F16    = 1,
     HTP_TYPE_Q4_0   = 2,
+    HTP_TYPE_Q4_1   = 3,
     HTP_TYPE_Q8_0   = 8,
     HTP_TYPE_IQ4_NL = 20,
     HTP_TYPE_I32    = 26,
@@ -28,6 +29,7 @@ enum htp_data_type {
 
     // types used internally for repack, dyn.quant, etc
     HTP_TYPE_Q4_0x4x2 = 200,
+    HTP_TYPE_Q4_1x4x2,
     HTP_TYPE_Q8_0x4x2,
     HTP_TYPE_MXFP4x4x2,
 
@@ -56,6 +58,7 @@ enum htp_op_code {
     HTP_OP_MUL_MAT,
     HTP_OP_MUL_MAT_ID,
     HTP_OP_RMS_NORM,
+    HTP_OP_RMS_NORM_MUL,
     HTP_OP_UNARY_SILU,
     HTP_OP_UNARY_GELU,
     HTP_OP_UNARY_SIGMOID,
@@ -89,6 +92,7 @@ enum htp_op_code {
     HTP_OP_TRI,
     HTP_OP_PAD,
     HTP_OP_NORM,
+    HTP_OP_CONCAT,
 
     HTP_OP_INVALID
 };

ggml/src/ggml-hexagon/htp/hvx-sin-cos.h

@@ -0,0 +1,90 @@
+#ifndef HVX_SIN_COS_H
+#define HVX_SIN_COS_H
+
+#include "hvx-base.h"
+#include "hvx-floor.h"
+
+static inline HVX_Vector hvx_vec_cos_f32(HVX_Vector x) {
+    HVX_Vector const_inv_pi = hvx_vec_splat_f32(0.3183098861837907f);
+    HVX_Vector const_half   = hvx_vec_splat_f32(0.5f);
+    HVX_Vector const_pi     = hvx_vec_splat_f32(3.141592653589793f);
+    HVX_Vector const_one    = hvx_vec_splat_f32(1.0f);
+    HVX_Vector const_neg_one = hvx_vec_splat_f32(-1.0f);
+
+    // n = floor(x * (1/pi) + 0.5)
+    HVX_Vector n_float = hvx_vec_floor_f32(hvx_vec_add_f32_f32(hvx_vec_mul_f32_f32(x, const_inv_pi), const_half));
+
+    // y = x - n * pi
+    HVX_Vector y = hvx_vec_sub_f32_f32(x, hvx_vec_mul_f32_f32(n_float, const_pi));
+
+    // Sign determination: if n is odd, sign is -1.0f, else 1.0f
+    // half_n = n * 0.5f
+    HVX_Vector half_n = hvx_vec_mul_f32_f32(n_float, const_half);
+    // floor_half_n = floor(half_n)
+    HVX_Vector floor_half_n = hvx_vec_floor_f32(half_n);
+    // is_odd = half_n > floor_half_n
+    HVX_VectorPred is_odd = Q6_Q_vcmp_gt_VsfVsf(half_n, floor_half_n);
+    // sign = vmux(is_odd, -1.0f, 1.0f)
+    HVX_Vector sign = Q6_V_vmux_QVV(is_odd, const_neg_one, const_one);
+
+    // z = y^2
+    HVX_Vector z = hvx_vec_mul_f32_f32(y, y);
+
+    // Chebyshev approximation for cos(y)
+    HVX_Vector c4 = hvx_vec_splat_f32(2.3557242013849433e-05f);
+    HVX_Vector c3 = hvx_vec_splat_f32(-0.0013871428263450528f);
+    HVX_Vector c2 = hvx_vec_splat_f32(0.041665895266688284f);
+    HVX_Vector c1 = hvx_vec_splat_f32(-0.4999999360426369f);
+    HVX_Vector c0 = hvx_vec_splat_f32(0.9999999999071725f);
+
+    HVX_Vector cos_y = hvx_vec_add_f32_f32(c3, hvx_vec_mul_f32_f32(z, c4));
+    cos_y = hvx_vec_add_f32_f32(c2, hvx_vec_mul_f32_f32(z, cos_y));
+    cos_y = hvx_vec_add_f32_f32(c1, hvx_vec_mul_f32_f32(z, cos_y));
+    cos_y = hvx_vec_add_f32_f32(c0, hvx_vec_mul_f32_f32(z, cos_y));
+
+    return hvx_vec_mul_f32_f32(cos_y, sign);
+}
+
+static inline HVX_Vector hvx_vec_sin_f32(HVX_Vector x) {
+    HVX_Vector const_inv_pi = hvx_vec_splat_f32(0.3183098861837907f);
+    HVX_Vector const_half   = hvx_vec_splat_f32(0.5f);
+    HVX_Vector const_pi     = hvx_vec_splat_f32(3.141592653589793f);
+    HVX_Vector const_one    = hvx_vec_splat_f32(1.0f);
+    HVX_Vector const_neg_one = hvx_vec_splat_f32(-1.0f);
+
+    // n = floor(x * (1/pi) + 0.5)
+    HVX_Vector n_float = hvx_vec_floor_f32(hvx_vec_add_f32_f32(hvx_vec_mul_f32_f32(x, const_inv_pi), const_half));
+
+    // y = x - n * pi
+    HVX_Vector y = hvx_vec_sub_f32_f32(x, hvx_vec_mul_f32_f32(n_float, const_pi));
+
+    // Sign determination: if n is odd, sign is -1.0f, else 1.0f
+    // half_n = n * 0.5f
+    HVX_Vector half_n = hvx_vec_mul_f32_f32(n_float, const_half);
+    // floor_half_n = floor(half_n)
+    HVX_Vector floor_half_n = hvx_vec_floor_f32(half_n);
+    // is_odd = half_n > floor_half_n
+    HVX_VectorPred is_odd = Q6_Q_vcmp_gt_VsfVsf(half_n, floor_half_n);
+    // sign = vmux(is_odd, -1.0f, 1.0f)
+    HVX_Vector sign = Q6_V_vmux_QVV(is_odd, const_neg_one, const_one);
+
+    // z = y^2
+    HVX_Vector z = hvx_vec_mul_f32_f32(y, y);
+
+    // Chebyshev approximation for sin(y)
+    HVX_Vector s4 = hvx_vec_splat_f32(2.642186986152672e-06f);
+    HVX_Vector s3 = hvx_vec_splat_f32(-0.00019825318964070864f);
+    HVX_Vector s2 = hvx_vec_splat_f32(0.00833326283319605f);
+    HVX_Vector s1 = hvx_vec_splat_f32(-0.16666666082087775f);
+    HVX_Vector s0 = hvx_vec_splat_f32(0.999999999915155f);
+
+    HVX_Vector sin_y = hvx_vec_add_f32_f32(s3, hvx_vec_mul_f32_f32(z, s4));
+    sin_y = hvx_vec_add_f32_f32(s2, hvx_vec_mul_f32_f32(z, sin_y));
+    sin_y = hvx_vec_add_f32_f32(s1, hvx_vec_mul_f32_f32(z, sin_y));
+    sin_y = hvx_vec_add_f32_f32(s0, hvx_vec_mul_f32_f32(z, sin_y));
+    sin_y = hvx_vec_mul_f32_f32(y, sin_y);
+
+    return hvx_vec_mul_f32_f32(sin_y, sign);
+}
+
+#endif /* HVX_SIN_COS_H */

ggml/src/ggml-hexagon/htp/hvx-utils.h

@@ -14,6 +14,8 @@
 #include "hvx-sqrt.h"
 #include "hvx-arith.h"
 #include "hvx-div.h"
+#include "hvx-floor.h"
+#include "hvx-sin-cos.h"
 #include "hvx-base.h"
 
 #endif /* HVX_UTILS_H */

ggml/src/ggml-hexagon/htp/main.c

@@ -420,8 +420,7 @@ AEEResult htp_iface_start(remote_handle64 handle, uint32 sess_id, uint64 dsp_que
 
     ctx->n_threads = n_hvx;
     for (int i = 0; i < ctx->n_threads; i++) {
-        // see discussion https://github.com/ggml-org/llama.cpp/pull/18151#discussion_r2632388541
-        ctx->dma[i] = dma_queue_create(128);
+        ctx->dma[i] = dma_queue_create(256); // queue depth
     }
 
     // init worker pool
@@ -538,6 +537,7 @@ static int execute_op(struct htp_ops_context * octx) {
 
         case HTP_OP_NORM:
         case HTP_OP_RMS_NORM:
+        case HTP_OP_RMS_NORM_MUL:
         case HTP_OP_SCALE:
         case HTP_OP_SQR:
         case HTP_OP_SQRT:
@@ -601,6 +601,9 @@ static int execute_op(struct htp_ops_context * octx) {
         case HTP_OP_PAD:
             return op_pad(octx);
 
+        case HTP_OP_CONCAT:
+            return op_concat(octx);
+
         case HTP_OP_GATED_DELTA_NET:
             return op_gated_delta_net(octx);
 
@@ -851,6 +854,11 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
         for (uint32_t i=0; i < n_ops; i++) {
             struct profile_data prof;
 
+            if (i == (n_ops-1)) {
+                // wake up the host before starting the last op
+                dspqueue_write_early_wakeup_noblock(queue, 0, 0);
+            }
+
             profile_start(ctx->profiler, &prof);
 
             proc_op_req(octx, tens, i, &ops[i]);
@@ -867,8 +875,6 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
             }
         }
 
-        // dspqueue_write_early_wakeup_noblock(ctx->queue, 10, 0);
-
         struct htp_opbatch_rsp rsp;
         rsp.id        = req.id;
         rsp.status    = HTP_STATUS_OK;

ggml/src/ggml-hexagon/htp/rope-ops.c

@@ -7,6 +7,7 @@
 
 #include <math.h>
 #include <string.h>
+#include <stdlib.h>
 
 #include "hex-dma.h"
 #include "hvx-utils.h"
@@ -75,6 +76,9 @@ struct htp_rope_context {
     size_t theta_cache_offset;
     uint32_t src0_nrows;
 
+    struct fastdiv_values div_ne2_ne1;
+    struct fastdiv_values div_ne1;
+
     uint64_t t_start;
 };
 
@@ -117,13 +121,84 @@ static __attribute__((noinline)) void rope_cache_init(const float    theta_base,
                             float *        cache,
                             const float    theta_scale) {
     // ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py
-    float theta = theta_base;
+#if __HVX_ARCH__ >= 79
+    const bool is_v79_or_newer = true;
+#else
+    const bool is_v79_or_newer = false;
+#endif
 
-    for (uint32_t i0 = 0; i0 < ne0; i0 += 2) {
-        const float ff = freq_factors ? freq_factors[i0 / 2] : 1.0f;
-        rope_yarn_one(theta / ff, freq_scale, corr_dims, i0, ext_factor, mscale, cache);
+    if (is_v79_or_newer && ext_factor == 0.0f) {
+        // Fast path: fully vectorized
+        // We process 32 pairs (64 elements) per iteration.
+        const uint32_t n_blocks = ne0 / 64;
 
-        theta *= theta_scale;
+        // Initialize theta scale powers: [1.0f, theta_scale, theta_scale^2, ..., theta_scale^31]
+        float __attribute__((aligned(128))) theta_powers[32];
+        theta_powers[0] = 1.0f;
+        for (int j = 1; j < 32; j++) {
+            theta_powers[j] = theta_powers[j - 1] * theta_scale;
+        }
+        HVX_Vector v_theta_powers = hvx_vmem(theta_powers);
+
+        HVX_Vector v_freq_scale = hvx_vec_splat_f32(freq_scale);
+        HVX_Vector v_mscale = hvx_vec_splat_f32(mscale);
+
+        // Base theta starts at theta_base
+        float theta_block = theta_base;
+        // The scale factor for the next block is theta_scale^32
+        float theta_scale_32 = 1.0f;
+        for (int j = 0; j < 32; j++) {
+            theta_scale_32 *= theta_scale;
+        }
+
+        for (uint32_t b = 0; b < n_blocks; b++) {
+            uint32_t i0 = b * 64;
+            HVX_Vector v_theta_base = hvx_vec_splat_f32(theta_block);
+            HVX_Vector v_theta = hvx_vec_mul_f32_f32(v_theta_base, v_theta_powers);
+
+            if (freq_factors) {
+                // Load 32 elements of freq_factors
+                HVX_Vector v_ff = hvx_vmemu(freq_factors + i0 / 2);
+                HVX_Vector v_inv_ff = hvx_vec_inverse_f32(v_ff);
+                v_theta = hvx_vec_mul_f32_f32(v_theta, v_inv_ff);
+            }
+
+            HVX_Vector v_theta_final = hvx_vec_mul_f32_f32(v_theta, v_freq_scale);
+
+            HVX_Vector vcos = hvx_vec_cos_f32(v_theta_final);
+            HVX_Vector vsin = hvx_vec_sin_f32(v_theta_final);
+
+            vcos = hvx_vec_mul_f32_f32(vcos, v_mscale);
+            vsin = hvx_vec_mul_f32_f32(vsin, v_mscale);
+
+            HVX_VectorPair vstore = Q6_W_vshuff_VVR(vsin, vcos, -4);
+
+            if (((uintptr_t)cache) % 128 == 0) {
+                hvx_vmem(cache + i0 + 0)  = Q6_V_lo_W(vstore);
+                hvx_vmem(cache + i0 + 32) = Q6_V_hi_W(vstore);
+            } else {
+                hvx_vec_store_u(cache + i0 + 0,  32 * sizeof(float), Q6_V_lo_W(vstore));
+                hvx_vec_store_u(cache + i0 + 32, 32 * sizeof(float), Q6_V_hi_W(vstore));
+            }
+
+            theta_block *= theta_scale_32;
+        }
+
+        // Leftovers
+        float theta = theta_block;
+        for (uint32_t i0 = n_blocks * 64; i0 < ne0; i0 += 2) {
+            const float ff = freq_factors ? freq_factors[i0 / 2] : 1.0f;
+            rope_yarn_one(theta / ff, freq_scale, corr_dims, i0, ext_factor, mscale, cache);
+            theta *= theta_scale;
+        }
+    } else {
+        // Fallback to original scalar loop
+        float theta = theta_base;
+        for (uint32_t i0 = 0; i0 < ne0; i0 += 2) {
+            const float ff = freq_factors ? freq_factors[i0 / 2] : 1.0f;
+            rope_yarn_one(theta / ff, freq_scale, corr_dims, i0, ext_factor, mscale, cache);
+            theta *= theta_scale;
+        }
     }
 }
 
@@ -195,24 +270,18 @@ static void rope_corr_dims(int     n_dims,
 }
 
 static inline void hvx_rope_neox_f32_aa(float * restrict dst, const float * restrict src0, uint32_t ne, const float * restrict theta_cache) {
-    const HVX_Vector * restrict vsrc   = (const HVX_Vector *) src0;
-    const HVX_Vector * restrict vtheta = (const HVX_Vector *) theta_cache;
-    HVX_Vector       * restrict vdst   = (HVX_Vector *) dst;
+    const uint32_t he = ne / 2;
+    const uint32_t nvec = he / 32;
+    const uint32_t nloe = he % 32;
 
-    uint32_t nvec = (ne / (VLEN_FP32 * 2) * 2); // 2 vecs per loop, step of 2
+    for (uint32_t i = 0; i < nvec; i++) {
+        HVX_Vector v0 = ((const HVX_Vector *) src0)[i];
+        HVX_Vector v1 = hvx_vmemu(src0 + he + i * 32);
 
-    uint32_t he = ne / 2;         // half_dims offset in elements
-    uint32_t hv = he / VLEN_FP32; // half_dims offset in vectors
+        HVX_Vector v2 = ((const HVX_Vector *) theta_cache)[i * 2 + 0];
+        HVX_Vector v3 = ((const HVX_Vector *) theta_cache)[i * 2 + 1];
 
-    #pragma unroll(2)
-    for (uint32_t i = 0; i < nvec; i += 2) {
-        HVX_Vector v0 = vsrc[i/2+0];
-        HVX_Vector v1 = vsrc[i/2+hv];
-
-        HVX_Vector v2 = vtheta[i+0];
-        HVX_Vector v3 = vtheta[i+1];
-
-        HVX_VectorPair vcos_sin = Q6_W_vdeal_VVR(v3, v2, -4);  // vcos_sin[0] = cos_theta, vcos_sin[1] = sin_theta
+        HVX_VectorPair vcos_sin = Q6_W_vdeal_VVR(v3, v2, -4);
 
         HVX_Vector vx0_c = Q6_Vqf32_vmpy_VsfVsf(v0, Q6_V_lo_W(vcos_sin));
         HVX_Vector vx0_s = Q6_Vqf32_vmpy_VsfVsf(v0, Q6_V_hi_W(vcos_sin));
@@ -222,37 +291,45 @@ static inline void hvx_rope_neox_f32_aa(float * restrict dst, const float * rest
         HVX_Vector v4 = Q6_Vqf32_vsub_Vqf32Vqf32(vx0_c, vx1_s);
         HVX_Vector v5 = Q6_Vqf32_vadd_Vqf32Vqf32(vx0_s, vx1_c);
 
-        vdst[i/2+0]  = Q6_Vsf_equals_Vqf32(v4);
-        vdst[i/2+hv] = Q6_Vsf_equals_Vqf32(v5);
+        ((HVX_Vector *) dst)[i] = Q6_Vsf_equals_Vqf32(v4);
+        hvx_vmemu(dst + he + i * 32) = Q6_Vsf_equals_Vqf32(v5);
     }
 
-    for (uint32_t i = nvec * VLEN_FP32; i < ne; i += 2) {
-        const float cos_theta = theta_cache[i+0];
-        const float sin_theta = theta_cache[i+1];
-        float x0 = src0[i/2];
-        float x1 = src0[i/2 + he];
-        dst[i/2]      = x0 * cos_theta - x1 * sin_theta;
-        dst[i/2 + he] = x0 * sin_theta + x1 * cos_theta;
+    if (nloe > 0) {
+        HVX_Vector v0 = hvx_vmemu(src0 + nvec * 32);
+        HVX_Vector v1 = hvx_vmemu(src0 + he + nvec * 32);
+
+        HVX_Vector v2 = ((const HVX_Vector *) theta_cache)[nvec * 2 + 0];
+        HVX_Vector v3 = ((const HVX_Vector *) theta_cache)[nvec * 2 + 1];
+
+        HVX_VectorPair vcos_sin = Q6_W_vdeal_VVR(v3, v2, -4);
+
+        HVX_Vector vx0_c = Q6_Vqf32_vmpy_VsfVsf(v0, Q6_V_lo_W(vcos_sin));
+        HVX_Vector vx0_s = Q6_Vqf32_vmpy_VsfVsf(v0, Q6_V_hi_W(vcos_sin));
+        HVX_Vector vx1_c = Q6_Vqf32_vmpy_VsfVsf(v1, Q6_V_lo_W(vcos_sin));
+        HVX_Vector vx1_s = Q6_Vqf32_vmpy_VsfVsf(v1, Q6_V_hi_W(vcos_sin));
+
+        HVX_Vector v4 = Q6_Vqf32_vsub_Vqf32Vqf32(vx0_c, vx1_s);
+        HVX_Vector v5 = Q6_Vqf32_vadd_Vqf32Vqf32(vx0_s, vx1_c);
+
+        hvx_vec_store_u(dst + nvec * 32, nloe * sizeof(float), Q6_Vsf_equals_Vqf32(v4));
+        hvx_vec_store_u(dst + he + nvec * 32, nloe * sizeof(float), Q6_Vsf_equals_Vqf32(v5));
     }
 }
 
 static inline void hvx_rope_f32_aa(float * restrict dst, const float * restrict src0, uint32_t ne, const float * restrict theta_cache) {
-    const HVX_Vector * restrict vsrc   = (const HVX_Vector *) src0;
-    const HVX_Vector * restrict vtheta = (const HVX_Vector *) theta_cache;
-    HVX_Vector       * restrict vdst   = (HVX_Vector *) dst;
+    const uint32_t nvec = ne / 64;
+    const uint32_t nloe = ne % 64;
 
-    uint32_t nvec = (ne / (VLEN_FP32 * 2)) * 2; // 2 vecs per loop, step of two
+    for (uint32_t i = 0; i < nvec; i++) {
+        HVX_Vector v0 = ((const HVX_Vector *) src0)[i * 2 + 0];
+        HVX_Vector v1 = ((const HVX_Vector *) src0)[i * 2 + 1];
 
-    #pragma unroll(2)
-    for (uint32_t i = 0; i < nvec; i+=2) {
-        HVX_Vector v0 = vsrc[i+0];
-        HVX_Vector v1 = vsrc[i+1];
+        HVX_Vector v2 = ((const HVX_Vector *) theta_cache)[i * 2 + 0];
+        HVX_Vector v3 = ((const HVX_Vector *) theta_cache)[i * 2 + 1];
 
-        HVX_Vector v2 = vtheta[i+0];
-        HVX_Vector v3 = vtheta[i+1];
-
-        HVX_VectorPair vx0_x1   = Q6_W_vdeal_VVR(v1, v0, -4);  // vx0_x1[0] = x0, vx0_x1[1] = x1
-        HVX_VectorPair vcos_sin = Q6_W_vdeal_VVR(v3, v2, -4);  // vcos_sin[0] = cos_theta, vcos_sin[1] = sin_theta
+        HVX_VectorPair vx0_x1   = Q6_W_vdeal_VVR(v1, v0, -4);
+        HVX_VectorPair vcos_sin = Q6_W_vdeal_VVR(v3, v2, -4);
 
         HVX_Vector vx0_c = Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(vx0_x1), Q6_V_lo_W(vcos_sin));
         HVX_Vector vx0_s = Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(vx0_x1), Q6_V_hi_W(vcos_sin));
@@ -264,17 +341,52 @@ static inline void hvx_rope_f32_aa(float * restrict dst, const float * restrict
 
         HVX_VectorPair vstore = Q6_W_vshuff_VVR(Q6_Vsf_equals_Vqf32(v5), Q6_Vsf_equals_Vqf32(v4), -4);
 
-        vdst[i+0] = Q6_V_lo_W(vstore);
-        vdst[i+1] = Q6_V_hi_W(vstore);
+        ((HVX_Vector *) dst)[i * 2 + 0] = Q6_V_lo_W(vstore);
+        ((HVX_Vector *) dst)[i * 2 + 1] = Q6_V_hi_W(vstore);
     }
 
-    for (uint32_t i = nvec * VLEN_FP32; i < ne; i += 2) {
-        const float cos_theta = theta_cache[i+0];
-        const float sin_theta = theta_cache[i+1];
-        float x0 = src0[i+0];
-        float x1 = src0[i+1];
-        dst[i+0] = x0 * cos_theta - x1 * sin_theta;
-        dst[i+1] = x0 * sin_theta + x1 * cos_theta;
+    if (nloe > 0) {
+        if (nloe <= 32) {
+            HVX_Vector v0 = hvx_vmemu(src0 + nvec * 64);
+            HVX_Vector v2 = hvx_vmemu(theta_cache + nvec * 64);
+
+            HVX_VectorPair vx0_x1   = Q6_W_vdeal_VVR(Q6_V_vzero(), v0, -4);
+            HVX_VectorPair vcos_sin = Q6_W_vdeal_VVR(Q6_V_vzero(), v2, -4);
+
+            HVX_Vector vx0_c = Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(vx0_x1), Q6_V_lo_W(vcos_sin));
+            HVX_Vector vx0_s = Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(vx0_x1), Q6_V_hi_W(vcos_sin));
+            HVX_Vector vx1_c = Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(vx0_x1), Q6_V_lo_W(vcos_sin));
+            HVX_Vector vx1_s = Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(vx0_x1), Q6_V_hi_W(vcos_sin));
+
+            HVX_Vector v4 = Q6_Vqf32_vsub_Vqf32Vqf32(vx0_c, vx1_s);
+            HVX_Vector v5 = Q6_Vqf32_vadd_Vqf32Vqf32(vx0_s, vx1_c);
+
+            HVX_VectorPair vstore = Q6_W_vshuff_VVR(Q6_Vsf_equals_Vqf32(v5), Q6_Vsf_equals_Vqf32(v4), -4);
+
+            hvx_vec_store_u(dst + nvec * 64, nloe * sizeof(float), Q6_V_lo_W(vstore));
+        } else {
+            HVX_Vector v0 = hvx_vmemu(src0 + nvec * 64);
+            HVX_Vector v1 = hvx_vmemu(src0 + nvec * 64 + 32);
+
+            HVX_Vector v2 = hvx_vmemu(theta_cache + nvec * 64);
+            HVX_Vector v3 = hvx_vmemu(theta_cache + nvec * 64 + 32);
+
+            HVX_VectorPair vx0_x1   = Q6_W_vdeal_VVR(v1, v0, -4);
+            HVX_VectorPair vcos_sin = Q6_W_vdeal_VVR(v3, v2, -4);
+
+            HVX_Vector vx0_c = Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(vx0_x1), Q6_V_lo_W(vcos_sin));
+            HVX_Vector vx0_s = Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(vx0_x1), Q6_V_hi_W(vcos_sin));
+            HVX_Vector vx1_c = Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(vx0_x1), Q6_V_lo_W(vcos_sin));
+            HVX_Vector vx1_s = Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(vx0_x1), Q6_V_hi_W(vcos_sin));
+
+            HVX_Vector v4 = Q6_Vqf32_vsub_Vqf32Vqf32(vx0_c, vx1_s);
+            HVX_Vector v5 = Q6_Vqf32_vadd_Vqf32Vqf32(vx0_s, vx1_c);
+
+            HVX_VectorPair vstore = Q6_W_vshuff_VVR(Q6_Vsf_equals_Vqf32(v5), Q6_Vsf_equals_Vqf32(v4), -4);
+
+            ((HVX_Vector *) dst)[nvec * 2 + 0] = Q6_V_lo_W(vstore);
+            hvx_vec_store_u(dst + nvec * 64 + 32, (nloe - 32) * sizeof(float), Q6_V_hi_W(vstore));
+        }
     }
 }
 
@@ -348,13 +460,19 @@ static void rope_job_f32(unsigned int nth, unsigned int ith, void * data) {
     const int32_t * pos = (const int32_t *) src1->data;
     const float * freq_factors = src2 ? (const float *) src2->data : NULL;
 
-    uint32_t ir = 0;
+    const uint32_t i3_start = fastdiv(src0_start_row, &rctx->div_ne2_ne1);
+    const uint32_t rem      = fastmodulo(src0_start_row, ne2 * ne1, &rctx->div_ne2_ne1);
+    const uint32_t i2_start = fastdiv(rem, &rctx->div_ne1);
+    const uint32_t i1_start = fastmodulo(rem, ne1, &rctx->div_ne1);
+
+    uint32_t ir = src0_start_row;
     uint32_t prev_i2 = (uint32_t) -1;
 
-    for (uint32_t i3 = 0; i3 < ne3; i3++) { // batch
-        for (uint32_t i2 = 0; i2 < ne2; i2++) { // seq-len
-            for (uint32_t i1 = 0; i1 < ne1; ) { // attn-heads
-                if (ir < src0_start_row) { ir++; i1++; continue; }
+    for (uint32_t i3 = i3_start; i3 < ne3; i3++) { // batch
+        const uint32_t i2_init = (i3 == i3_start) ? i2_start : 0;
+        for (uint32_t i2 = i2_init; i2 < ne2; i2++) { // seq-len
+            const uint32_t i1_init = (i3 == i3_start && i2 == i2_start) ? i1_start : 0;
+            for (uint32_t i1 = i1_init; i1 < ne1; ) { // attn-heads
                 if (ir >= src0_end_row) goto done;
 
                 // Rows in this block
@@ -407,9 +525,6 @@ static void rope_job_f32(unsigned int nth, unsigned int ith, void * data) {
                                         ne0, rctx->ext_factor, rctx->attn_factor,
                                         theta_cache, rctx->theta_scale);
                     }
-
-                    // FARF(HIGH, "rope-theta %u: ir %u i1 %u i2 %u i3 %u cache %p : usec %u", ith, ir, i1, i2, i3, theta_cache,
-                    //         (unsigned) HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - rctx->t_start));
                 }
 
                 // Skip output DMA transactions from prev block (if any)
@@ -489,7 +604,7 @@ static int execute_op_rope_f32(struct htp_ops_context * octx) {
     // Aligned row sizes for VTCM
     const size_t src0_row_size_aligned    = hex_round_up(src0_row_size, VLEN);
     const size_t dst_row_size_aligned     = hex_round_up(dst_row_size, VLEN);
-    const size_t theta_cache_size_aligned = hex_round_up(src0->ne[0] * sizeof(float), 128);
+    const size_t theta_cache_size_aligned = hex_round_up(src0->ne[0] * sizeof(float), 256);
 
     // Calculate spad sizes per thread
     size_t src0_spad_per_thread = theta_cache_size_aligned + HTP_ROPE_SPAD_NROWS * src0_row_size_aligned;
@@ -546,6 +661,11 @@ static int execute_op_rope_f32(struct htp_ops_context * octx) {
     rctx.src0_nrows = src0_nrows;
     rctx.src0_nrows_per_thread = (src0_nrows + n_threads - 1) / n_threads;
 
+    if (src0_nrows > 0) {
+        rctx.div_ne2_ne1 = init_fastdiv_values(dst->ne[2] * dst->ne[1]);
+        rctx.div_ne1     = init_fastdiv_values(dst->ne[1]);
+    }
+
     FARF(HIGH, "rope-f32 n-rows %u n-dims %d ne0 %u ext-factor %.6f theta-scale %.6f attn-factor %.6f\n", rctx.src0_nrows, rctx.n_dims, ne0,
          rctx.ext_factor, rctx.theta_scale, rctx.attn_factor);
 

ggml/src/ggml-hexagon/htp/set-rows-ops.c

@@ -65,6 +65,9 @@ static void set_rows_thread_f32_f32(unsigned int nth, unsigned int ith, void *da
     // parallelize by rows of src0
     const uint32_t dr  = srctx->src0_nrows_per_thread;
     const uint32_t ir0 = dr * ith;
+    if (ir0 >= nr) {
+        return;
+    }
     const uint32_t ir1 = (ir0 + dr < nr) ? (ir0 + dr) : nr;
 
     const bool is_i32 = (octx->src[1]->type == HTP_TYPE_I32);
@@ -109,6 +112,9 @@ static void set_rows_thread_f16_f32(unsigned int nth, unsigned int ith, void *da
     // parallelize by rows of src0
     const uint32_t dr  = srctx->src0_nrows_per_thread;
     const uint32_t ir0 = dr * ith;
+    if (ir0 >= nr) {
+        return;
+    }
     const uint32_t ir1 = (ir0 + dr < nr) ? (ir0 + dr) : nr;
 
     const bool is_i32 = (octx->src[1]->type == HTP_TYPE_I32);